Akash Systems uses diamond-based cooling technology that eliminates thermal constraints, unlocking higher performance, greater efficiency, longer system life, and lower costs. By bonding synthetic diamond to chip surfaces, they pull heat away far more effectively than traditional methods, allowing devices to run faster and last longer.
Akash Systems' Diamond Cooling fuses the superior thermal conductivity of synthetic diamond materials with the semiconductors that power modern electronics, creating a new material that dissipates heat efficiently, boosts energy efficiency, enhances system-wide performance, and increases the lifespan of electronics.
The massive heat generated by AI training and inference workloads on GPUs has created urgent demand for Akash's Diamond Cooling server technology, which enables 25% overclocking and doubles server lifetimes by eliminating thermal throttling.
Atomic Machines is building a new type of robotic manufacturing system that can create tiny mechanical devices with features just a few millionths of a meter in size. Their long-term goal is to make manufacturing so precise that objects can be assembled at the atomic level, starting with components for the AI data center market.
Unlike conventional MEMS fabrication tied to silicon fab limitations, Atomic Machines' robotic platform enables rapid prototyping and large-scale production of complex 3D microdevices with single-digit-micron features. Their approach breaks traditional manufacturing constraints, allowing novel device geometries and materials combinations impossible with existing methods.
Atomic Machines has developed its first device targeting the AI data center market, capitalizing on the need for novel micro-components that can improve the performance and efficiency of AI infrastructure. Atomic Machines has raised $144M in total funding backed by investors including Sozo Ventures and XTX Ventures, and received a major California Competes tax credit as part of a $156.3M investment to expand its MEMS fabrication facilities in Santa Clara and Emeryville.
Ayar Labs replaces the traditional copper wire connections between chips with tiny light-based links that can move data much faster while using less power. Their optical connectors fit inside standard chip packages, making it possible to dramatically speed up data movement without overhauling existing manufacturing processes.
While other optical interconnect companies focus on longer-distance fiber optic links, Ayar Labs brings optics inside the chip package itself, addressing the data movement bottleneck at its most fundamental level. Their technology is compatible with existing semiconductor manufacturing processes, enabling adoption without requiring entirely new fab infrastructure.
AI training clusters require moving enormous amounts of data between GPUs, CPUs, and memory at speeds that electrical interconnects increasingly cannot deliver, making Ayar Labs' optical I/O a critical enabling technology for next-generation AI systems. Ayar Labs has deepened partnerships with major chipmakers and received investment from strategic players including Nvidia, as demand for its optical I/O technology accelerates alongside AI infrastructure buildouts.
Baseten provides a cloud platform that makes it easy for companies to deploy and run AI models in production without managing the underlying server infrastructure. Their system automatically handles things like assigning the right number of graphics processors and scaling up or down based on demand.
Unlike general-purpose cloud providers, Baseten is built specifically for ML model deployment, offering optimized model serving with features like automatic batching, streaming responses, and built-in GPU scheduling. Their developer-first approach provides simple APIs and Truss, an open-source model packaging framework, that makes deployment drastically simpler than DIY alternatives.
The explosion of companies deploying AI models — from LLMs to image generators to recommendation systems — has created massive demand for inference infrastructure, fueling Baseten's rapid growth as a specialized platform. Baseten reached a $5B valuation with a $300M funding round led by IVP and CapitalG in January 2026.
ChipAgents builds AI agents that help engineers design and verify computer chips dramatically faster — think of it as a team of AI coworkers that can read chip specifications, write and debug design code, and run verification tests autonomously. Their platform plugs into engineers' existing code editors and workflows, aiming to make chip design 10 times more productive.
Unlike traditional chip design tools that assist with individual tasks, ChipAgents deploys coordinated teams of AI agents that take ownership of entire workflows — reading specs, breaking down objectives, implementing solutions, validating results, and iterating without constant human guidance. Their agents have demonstrated faster test generation and faster specification comprehension in production environments at major semiconductor companies.
The breakthroughs in agentic AI — where AI systems can autonomously plan, reason, and execute multi-step tasks — have made it possible for the first time to automate the complex, judgment-heavy work of chip design and verification. ChipAgents raised $74M in total funding including an oversubscribed $50M Series A1 in February 2026 led by TSMC-backed Matter Venture Partners, with Bessemer, Micron, MediaTek, and Ericsson also investing. The company reported 140x year-over-year revenue growth, deployments at 80 semiconductor companies, and advisory board additions including former CEOs of Cadence, Mentor Graphics (Siemens), and the former CTO of Synopsys.
Cognichip is building an AI system that can help engineers design computer chips faster and cheaper. Their goal is to cut chip development time in half and reduce costs by 75%, making it possible for more companies to afford designing their own custom chips.
While existing EDA tools assist with specific design tasks, Cognichip's ACI model acts like an expert chip engineer, understanding and solving design problems with designer-level cognitive abilities across the entire development process. The platform promises faster design cycles and cost reduction through concurrent local and global optimization, unlike the serial approach of traditional tools.
The generative AI revolution has made it possible for the first time to build foundational models that understand semiconductor physics deeply enough to automate and accelerate chip design at scale. Cognichip emerged from stealth in May 2025 with $33M in seed funding co-led by Lux Capital and Mayfield, assembling a team of AI experts from Stanford, Google, MIT, and semiconductor veterans from Amazon, Apple, and Synopsys.
Corintis builds tiny liquid cooling channels that sit directly on or near a chip's surface, pulling heat away right at the source rather than relying on fans or larger cooling systems further away. This approach is becoming essential as AI chips push past power levels that traditional cooling methods can no longer handle.
Traditional cooling solutions work at the system level, but Corintis' microfluidic technology operates at the chip level, channeling coolant through microscale structures directly on or near the silicon die. This removes heat much closer to its source than conventional approaches, enabling chips to run at higher power levels without thermal throttling.
Corintis raised $24M in Series A funding in late 2025, adding Intel CEO Lip Bu Tan to their board. They partnered with Microsoft to show 3x better cooling with microfluidics versus cooling alternatives.
Cornelis Networks makes high-speed networking equipment designed specifically for connecting thousands of AI processors together in a computing cluster. Their technology is optimized for the unique communication patterns of AI training, where processors constantly need to share data with each other at extremely low latency.
Cornelis' networking fabric is purpose-built for the collective communication patterns of AI training workloads, where thousands of GPUs must synchronize gradients and share data in patterns very different from traditional data center traffic. Their architecture delivers lower latency and higher message rate than general-purpose Ethernet for AI-specific communication. The company formed as a spinout of Intel's OmniPath networking product.
Crusoe builds AI data centers near renewable and stranded energy sources, reducing both cost and carbon footprint compared to conventional cloud providers. Their first product to market used stranded energy, like natural gas being burned off at oil wells or renewable energy that the grid can't absorb, to power colocated data centers.
The company operates H100 and H200 GPU clusters and is a preferred provider for major AI workloads, including a 2.1GW data center in Abilene, TX to support Microsoft.
The enormous and growing power demands of AI training and inference have made cheap energy access the key competitive advantage in AI infrastructure, perfectly aligning with Crusoe's stranded energy model. Crusoe raised a $1.3B Series E, led by Valor Equity Partners and Mubadala Capital in October 2025.
* Primary is an investor in Crusoe, through the acquisition of Atero.
d-Matrix builds processors that perform AI calculations directly inside the memory where data is stored, rather than constantly shuttling data back and forth between separate memory and processing units. This approach dramatically reduces the energy wasted on moving data around, which is one of the biggest inefficiencies in running AI today.
Unlike GPUs that shuttle data between separate compute and memory units, d-Matrix performs calculations directly inside memory arrays using a digital (not analog) approach, maintaining precision while dramatically reducing data movement energy. This gives them both the efficiency benefits of in-memory computing and the reliability of digital circuits.
As AI inference workloads grow to dominate data center spending — projected to surpass training costs — d-Matrix's inference-optimized architecture offers compelling total cost of ownership advantages over GPU-based solutions. d-Matrix raised a $275M Series C in Q4 2025 led by BullhoundCapital, Triatomic Capital, and Temasek, with participation from Microsoft's M12 and the Qatar Investment Authority.

DriveNets sells networking as software. Instead of buying proprietary routers, a carrier or AI datacenter operator runs DriveNets' software across a cluster of standard white-box switches that then behaves as one very large router, the same disaggregation that separated servers from mainframes, applied to the network.
Unlike Nvidia's InfiniBand and NVLink, which deliver high performance inside a single-vendor stack, DriveNets builds its AI fabric on standard Ethernet and positions itself as neutral infrastructure for mixed-vendor clusters, publishing open reference designs with AMD and Broadcom. Its fabric supports scale-up, scale-out and scale-across topologies alongside front-end and storage connectivity, and the pitch is measured in GPU utilization: higher tokens per second and lower cost per token from optimizing the stack end to end.
As operators build clusters spanning multiple accelerator vendors, the networking layer becomes the place where heterogeneity is either enabled or blocked, which is the wedge DriveNets is pressing. The telecom business gives it an unusual profile for a company in this category: cash-flow positive since 2025, with more than $1B in secured business and roughly $250M in trailing revenue. DriveNets closed a $410M Series D on 1 June 2026 led by Bessemer Venture Partners and Atreides Management, with AMD and Red Dot Capital joining as new investors alongside Pitango and D1 Capital, at an $8.5B post-money valuation and $1B raised to date.
Efficient Computer designs ultra-low-power general-purpose processors that run AI workloads on devices that can't be plugged into a wall — robots, sensors, satellites, and implantables — without draining their batteries. Their Fabric architecture, a spatial dataflow design developed at Carnegie Mellon University, eliminates the energy overhead of conventional von Neumann processors and delivers up to 100x better efficiency than traditional CPUs, while remaining fully programmable in standard languages.
Unlike fixed-function accelerators that trade flexibility for efficiency, Efficient's approach achieves dramatic power reduction without locking developers into specialized hardware or custom programming models. This makes it practical to run evolving AI software on energy-constrained devices without redesigning the chip every time models change.
Efficient Computer raised $76M in total funding across a $16M seed and a $60M Series A led by Triatomic Capital, with participation from Eclipse, Union Square Ventures, RTX Ventures, and Toyota Ventures.
Eliyan makes high-speed connection technology that lets multiple small chips inside a processor package communicate with each other as if they were a single large chip. Their approach works with standard, affordable packaging materials rather than requiring expensive specialty components, making chiplet-based designs more practical.
Unlike silicon interposers or bridges that are expensive and supply-constrained, Eliyan's interconnect works on standard organic packaging substrates while still achieving the bandwidth density typically associated with advanced packaging. This means chiplet-based designs can achieve high performance without the cost premium of exotic packaging technologies.
The shift to chiplet-based AI processor designs — combining specialized compute, memory, and networking dies — is creating strong demand for Eliyan's high-bandwidth interconnect technology that makes chiplet integration practical and affordable. Eliyan raised $62M in Series B funding in 2025 and announced key partnerships with semiconductor manufacturers, as its NuLink technology gained traction as a critical enabler of next-generation chiplet architectures.
Eridu is building an AI networking solution, designed from the ground up for scale-out and scale-up networks in AI data centers. Eridu’s network switch delivers the order-of-magnitude advance in performance, radix and throughput needed to maximize utilization, throughput and efficiency in AI clusters and data centers.
Demands for data movement are growing exponentially as hyperscalers, frontier labs and neoclouds build ever-larger AI data centers, exposing fragility in the networking stack that exists today. At the same time, new architectural and algorithmic innovations in AI are only making the network bottleneck worse.
Eridu raised an oversubscribed $200M Series A in March 2026, led by Socratic Partners, John Doerr, Hudson River Trading, Capricorn Investment Group and Matter Venture Partners, with participation by Bosch Ventures, Eclipse Capital, Fusion Fund, Osage University Partners, SBVA, TDK Ventures, VentureTech Alliance, and Zelda Ventures among others.
Etched is a Cupertino-based semiconductor startup, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen. The company is building frontier inference clusters. Their rack-scale systems leverage two core architectures, Low Voltage Inference and Cluster Scale Memory, so frontier models can run performantly across both prefill and decode workloads.
Etched has raised $800M in funding, with investors including Stripes, Ribbit Capital, Positive Sum, and Primary Venture Partners.
* Primary is an investor in Etched
Fireworks AI runs a platform where developers can access, customize, and deploy open-source AI models through a fast and affordable cloud service. They have deeply optimized how models run on graphics processors, squeezing significantly more speed out of each chip than standard approaches.
Fireworks differentiates through its compound AI system approach, allowing developers to compose multiple models and tools together in a single API call, and through deep kernel-level optimizations that squeeze more performance from each GPU. Their FireAttention engine and custom CUDA kernels deliver significantly faster token generation than standard serving frameworks.
The rapid proliferation of open-source AI models and enterprise adoption of LLMs has created a massive market for optimized inference platforms, driving Fireworks' growth as developers seek faster, cheaper alternatives to self-hosting. Fireworks AI raised $250M at a $4B valuation in October 2025 in a Series C led by Lightspeed Venture Partners, Index Ventures and Evantic.
Firmus Technologies builds and operates data centers designed from the ground up to handle AI workloads, with advanced power and cooling systems optimized for high-performance computing. Their facilities are purpose-built for AI rather than being traditional data centers retrofitted with graphics processors.
Firmus emphasizes operational efficiency and energy sustainability in its data center design, incorporating advanced power and cooling technologies to minimize environmental impact while maximizing compute density. Their facilities are designed from the ground up for AI workloads rather than being retrofitted from general-purpose data centers.
The surge in demand for AI training and inference capacity has created a massive need for purpose-built AI data centers, positioning Firmus as part of the new wave of AI-native infrastructure providers. Firmus raised nearly $550M in capital from Nvidia and others in September 2025. Soon after, they raised $10B in debt in January 2026 to fund Project Southgate, aimed to deploy infrastructure for sovereign AI in Australia.
FluidStack runs a distributed cloud that pulls together GPUs from data centers around the world, giving AI developers on-demand access to computing power at competitive prices. Their marketplace model aggregates capacity from multiple partners, often making hard-to-find AI chips more available than on the big cloud platforms. Unlike traditional cloud providers with centralized data centers, FluidStack's distributed model aggregates GPU capacity from multiple partners, enabling more competitive pricing and greater availability of hard-to-find AI accelerators. Their marketplace approach creates price competition among GPU suppliers, driving costs down for end users.
The chronic shortage of GPU compute for AI training has driven developers and companies to seek alternative cloud providers like FluidStack that can aggregate distributed capacity and offer better availability than hyperscalers.
FluidStack raised $700M in December 2025 from Situational Awareness, valuing the company at $7B. They are reportedly in talks to raise at an $18B valuation, as of April 2026.
Fractile is a UK-based startup designing chips that run large language models more efficiently by rethinking how numbers are represented and processed during AI calculations. Their approach squeezes more useful computation out of every watt of power, targeting the growing cost of running AI at scale. Fractile's architecture exploits the mathematical properties of LLMs to represent computations more efficiently than conventional floating-point approaches, extracting more useful work per watt than GPU-based solutions. Their approach targets the specific mathematical operations that dominate transformer inference, rather than trying to be a general-purpose accelerator.
The rapid scaling of LLM deployment across enterprise and consumer applications is creating massive demand for more efficient inference hardware, which Fractile's numerically optimized architecture directly addresses.
Fractile announced it will be investing over $100M into building out semiconductor capacity in the UK.
Gimlet Labs is building what it calls the first "multi-silicon inference cloud" — software that allows AI workloads to run simultaneously across diverse hardware types including CPUs, GPUs, and specialized accelerators. Its platform automatically disaggregates agentic workloads into component stages and maps each to the most suitable chip without requiring code changes. Rather than optimizing for a single chip vendor, Gimlet treats the entire heterogeneous data center fleet as one unified compute pool — addressing the problem that existing hardware runs at only 15 to 30 percent utilization. It partners across Nvidia, AMD, Intel, ARM, Cerebras, and d-Matrix, positioning itself as the VMware of AI inference rather than a chip-specific stack.
As AI moves from experimentation to production deployment across enterprises, the need for robust operational infrastructure and tooling is growing rapidly, directly driving demand for Gimlet Labs' solutions.
Gimlet raised an $80M Series A led by Menlo Ventures, bringing total funding to $92M, with angels including Intel CEO Lip-Bu Tan and former VMware CEO Raghu Raghuram. The company launched publicly in October 2025 with eight-figure revenues out of the gate and has since more than doubled its customer base across frontier AI labs and major cloud providers.
Lambda provides cloud computing and hardware specifically built for AI researchers and developers, offering ready-to-use servers loaded with the latest graphics processors and pre-configured AI software. They have become a go-to choice for machine learning teams that want powerful computing without the setup complexity of traditional cloud providers. Lambda is laser-focused on the ML/AI community rather than being a general cloud provider, offering pre-configured software stacks and developer-friendly tooling that eliminate the setup friction researchers face on traditional clouds. Their 1-Click Clusters feature enables teams to spin up multi-node GPU training infrastructure in minutes rather than weeks.
The explosive growth of AI research and commercial model development has driven surging demand for Lambda's ML-optimized cloud and hardware products, establishing it as the go-to GPU cloud for AI practitioners.
In November 2025, they raised $1.6B from TWG Global in a Series E funding round. As of January 2026, Lambda Labs was in talks to raise $350M in pre-IPO funding.
Lightmatter builds networking equipment that uses light instead of electricity to connect AI chips across a data center, delivering dramatically more bandwidth with less energy. Their technology can link chips together at speeds previously only possible within a single processor, removing the networking bottleneck that slows down large AI systems. Lightmatter's Passage interconnect is capable of connecting chips across an entire data center with the bandwidth density previously only possible within a single chip. Their programmable photonic technology can dynamically reconfigure data pathways, unlike fixed electrical connections.
As AI clusters scale to hundreds of thousands of GPUs, the network connecting them becomes the primary bottleneck; Lightmatter's photonic fabric directly solves this by providing orders of magnitude more bandwidth at lower power.
Lightmatter raised $400M in 2025, reaching a $4.4B valuation, and has been deepening partnerships with major AI chip and cloud companies to integrate its photonic interconnect technology into next-generation data center architectures.
.png)
Majestic Labs builds AI servers with custom accelerator and memory interface chips that disaggregate memory from compute, enabling up to 128 TB of high-bandwidth memory per server. Each server is designed to collapse multiple racks of conventional GPU equipment into a single unit for training and inference workloads. Majestic attacks the "memory wall" directly by delivering roughly 1,000x the memory capacity of a standard GPU server. The architecture is fully vertically integrated (custom silicon, systems, software) and maintains full programmability to support workloads beyond current transformer models.
The insatiable demand for more efficient AI computing has created opportunities for companies like Majestic Labs that bring fresh architectural thinking to the fundamental challenges of AI hardware design.
Majestic has raised over $100M in Series A financing and was founded by the team behind Meta's FAST silicon group and Google's GChips, holding 120+ patents collectively. The company is targeting hyperscalers and enterprises in data-intensive verticals like financial services and pharma.
MatX is a San Francisco-based AI chip company, founded in 2023, designing high-throughput accelerators specifically optimized for running large language models. The company went from a $25M seed round to over $600M in total funding in under two years. Their first chip, MatX One, is based on a splittable systolic array which combines the low latency of SRAM-first designs with the long-context support of HBM.
The MatX founding team comes from Google's TPU team. MatX raised a $500M round in 2026 and is building chips designed to deliver more compute per dollar for LLM inference than general-purpose GPUs.
Modal is a cloud platform that lets developers run AI and data-intensive code on powerful remote servers using nothing more than a few lines of Python — no infrastructure setup required. The platform instantly provisions the right hardware, including graphics processors, and scales automatically. Modal's key innovation is making cloud GPU access feel like running code locally — developers write normal Python functions and Modal handles everything from containerization to GPU allocation in milliseconds. Unlike traditional cloud providers that require provisioning VMs and managing Kubernetes, Modal's serverless approach means zero infrastructure management.
The AI boom has created millions of developers who need GPU access for model training, fine-tuning, and inference but lack infrastructure expertise, making Modal's frictionless cloud platform increasingly essential.
Modal announced a $366M Series C in May 2026, reaching a $4.65B valuation, as its sandboxes product has taken off, pushing the business past $300M ARR.

Nearfield Instruments builds machines that inspect computer chips while they are being made. Its systems drag an atomically sharp probe across the surface of a wafer to map features just billionths of a meter across, like a record needle fine enough to feel the shape of individual transistor structures, reading the wafer without damaging it.
Unlike conventional inspection, which often requires slicing a wafer apart to see inside it, destroying a product worth tens of thousands of dollars to learn whether the process is drifting, Nearfield's QUADRA system measures in full 3D non-destructively and in-line. The company's advance is throughput: an array of miniaturized atomic force microscope heads scanning in parallel, which it claims delivers over 100x the image acquisition speed of conventional single-probe automated AFM tools. Its AUDIRA system extends the same approach to buried features and subsurface voids, which matter increasingly as chips are stacked rather than shrunk.
The structures that AI chips depend on, including high-aspect-ratio DRAM and HBM stacks, gate-all-around and CFET logic, hybrid-bonded 3D packaging, and high-NA EUV resist layers, are precisely the ones conventional metrology struggles to measure, making process control a live constraint on advanced-node yield. QUADRA is qualified and deployed for high-volume manufacturing at leading fabs, and in November 2025 Nearfield signed a multi-year development agreement to install QUADRA at imec's R&D facility in Leuven. In February 2026 the company added a Sidewall Imaging Mode to QUADRA for measuring the flanks of high-aspect-ratio 3D logic and memory structures, unveiled at SPIE Advanced Lithography. The company closed an oversubscribed $380M Series D led by Fidelity in June 2026 at a $1.6B valuation, with Qatar Investment Authority joining alongside Temasek, Walden Catalyst, Innovation Industries, M&G, Invest-NL, TNO Ventures and ING.
Neurophos is developing chips that combine light-based computing elements with traditional electronic circuits on the same piece of silicon. This hybrid approach aims to deliver the speed and energy advantages of computing with light while remaining compatible with existing chip manufacturing infrastructure. Neurophos' approach integrates photonic computing elements with conventional electronic circuits on the same chip, offering a practical path to adopting optical computing without requiring a complete overhaul of existing infrastructure. Their hybrid approach balances the speed of light-based processing with the flexibility of electronics.
The energy and performance demands of AI inference and training are increasingly exceeding what purely electronic chips can deliver, creating a compelling use case for Neurophos' photonic computing technology.
In January 2026, Neurophos raised a $110M Series A, led by Gates Frontier, with participation from M12, Carbon Direct Capital, Silicon Catalyst Ventures, and more.
Nexthop AI builds custom networking equipment — switches, software, and optical connections — designed specifically for the massive AI data centers operated by the world's largest cloud companies. They have a unique co-development model, focused on delivering hardware and software purpose-built cloud providers running open source operating systems. Unlike traditional networking vendors that sell one-size-fits-all products, Nexthop co-engineers custom solutions alongside each hyperscaler's own team, integrating seamlessly into their specific technology stack. CEO Anshul Sadana previously spent 17 years at Arista Networks, serving as its Chief Customer Officer, then Chief Operating Officer.
AI training clusters require thousands of processors to exchange data at blinding speeds, making the network fabric between them a critical performance bottleneck — and Nexthop's purpose-built AI networking equipment directly addresses this growing need.
Nexthop AI raised an oversubscribed $500M Series B in March 2026 led by Lightspeed Venture Partners with Andreessen Horowitz joining, catapulting the company's valuation to $4.2 billion — just one year after emerging from stealth with $110M in funding. The company simultaneously unveiled three new Ethernet switch platforms engineered for hyperscale AI clusters and neoclouds.
Nscale is a European cloud provider building AI computing infrastructure across Europe, giving companies and governments access to powerful graphics processors while keeping their data on European soil. This addresses growing demand from organizations that need AI computing power but must comply with European data privacy regulations. Nscale positions itself as Europe's answer to U.S.-dominated GPU cloud providers, offering data sovereignty and compliance with EU regulations like GDPR for organizations that cannot or prefer not to send data to American hyperscalers. Their European-first approach addresses a critical gap in the market for sovereign AI infrastructure.
European companies and governments increasingly need local, sovereign AI compute infrastructure to comply with data regulations and reduce dependence on U.S. cloud providers, directly driving Nscale's growth.
In March 2026, Nscale and Microsoft announced a collaboration to to deliver 1.35GW of Vera Rubin NVIDIA capacity on a new West Virginia Campus.

OLIX Computing is building a new class of AI accelerator — the Optical Tensor Processing Unit (OTPU) — that uses photonics to run AI inference workloads faster and more cost-efficiently than GPU-based systems. By integrating SRAM memory directly with optical components, their architecture sidesteps the memory bandwidth bottleneck that limits conventional AI chips without relying on HBM, advanced packaging, or any supply-chain-constrained technology that even the largest hyperscalers struggle to procure.
Unlike GPU designs that depend on expensive high-bandwidth memory stacked alongside the chip, OLIX stores data entirely in on-chip SRAM and uses photonics for interconnect, achieving lower latency and higher throughput per watt at a significantly lower total cost of ownership. Their chips are designed for bit-perfect digital computation — not analog approximation — and are compatible with existing AI models through a custom compiler stack.
Founded in 2024 by 25-year-old James Dacombe, OLIX raised $250M in total funding — including a $220M round led by Hummingbird Ventures at a $1B+ valuation — and has grown to over 70 employees across the UK and North America. The company plans to ship its first OTPU products to customers in 2027.

Positron builds hardware dedicated to AI inference — the everyday work of answering model queries, rather than the one-time work of training. Its shipping product, Atlas, is a rack-mounted inference appliance that the company says delivers roughly three times the compute per watt of an Nvidia H100 and beats an H200 while drawing about a third of the power. Atlas is fabricated and assembled in the US, and serves requests through an OpenAI-compatible endpoint so customers can swap it in without rewriting their stack.
Unlike accelerators that compete on peak floating-point throughput, Positron designs around memory as the binding constraint. Its next-generation Asimov silicon is specified at up to 2.3TB of memory per chip against 384GB for Nvidia's upcoming Rubin, and the four-chip Titan system is designed to hold 8TB+ of memory, 10M-token context windows, and models up to 16 trillion parameters on a single server. That bet targets long-context LLMs, agentic workflows, and video models — workloads where capacity, not compute, runs out first. Atlas today is FPGA-based; Asimov is the move to custom silicon.
Inference has overtaken training as the dominant share of AI compute spend, and grid power has become the practical ceiling on deployment, making tokens-per-watt the metric customers actually buy on. Positron raised an $875M Series C in September 2026 at a $5B post-money valuation, structured as a $375M Series C co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital and SemiAnalysis Capital, plus a Series C-1 of up to $500M anchored by NEA and Netscape co-founder Jim Clark; NEA's Forest Baskett, Atreides' Gavin Baker, Thomas Jermoluk and Dylan Patel joined the board. The round followed a $230M Series B in February 2026 at a $1B+ valuation co-led by ARENA Private Wealth, Jump Trading and Unless, with Qatar Investment Authority and Arm participating, and brings total funding to roughly $1.18B. More than 50 Atlas racks are deployed, including at Oracle Cloud Infrastructure, serving customers such as Parasail, Jump Trading and i3d.net. Proceeds fund the Asimov tape-out on TSMC N3P by end of 2026, a 2MW+ engineering data center, and the production ramp of Titan, with Asimov production expected in the second half of 2027. CEO Mitesh Agrawal was previously COO of Lambda; co-founder and CTO Thomas Sohmers previously founded REX Computing.
PowerLattice makes a small chip component called a power delivery chiplet that sits inside a processor's package and delivers electricity right next to where the computing happens. By moving power regulation closer to the chip, they cut energy waste by more than half — a big deal when AI chips are consuming record amounts of electricity. While competitors deliver power from the motherboard inches away from the chip, PowerLattice places voltage regulation just hundreds of micrometers from the processor die, cutting compute power needs by more than 50%. Their chiplet-based approach is inherently scalable and can be deployed in parallel, adapting to any SoC power topology.
AI accelerators are straining data center power and cooling infrastructure, making PowerLattice's 50%+ power reduction directly critical to sustaining AI compute scaling.
PowerLattice emerged from stealth in November 2025 with a $25M Series A led by Playground Global and Celesta Capital (with former Intel CEO Pat Gelsinger as a board member).
PsiQuantum is building a large-scale quantum computer that uses particles of light instead of the electrical circuits used in traditional computers. Their key advantage is that they manufacture their quantum components in standard chip factories, making their approach uniquely scalable compared to other quantum products. Unlike competitors using superconducting or trapped-ion qubits that require exotic manufacturing, PsiQuantum fabricates its quantum components in standard silicon foundries, making its approach uniquely scalable to the millions of qubits needed for useful computation. Their photonic architecture operates at room temperature components where possible, avoiding the extreme cooling requirements of superconducting approaches.
Quantum computing promises to solve optimization and simulation problems that are intractable for classical AI, and PsiQuantum's photonic approach could enable quantum-enhanced machine learning algorithms that dramatically accelerate AI capabilities.
PsiQuantum raised $1.5B in a Series E at a $10.5B valuation, with investors including Nvidia, Blackrock, and Baillie Gifford.
Rebellions is a South Korean chipmaker designing processors built specifically for running AI workloads in both data centers and smaller devices at the network edge. As the leading AI chip company in South Korea, Rebellions benefits from strong government backing and partnerships with Samsung Foundry for manufacturing, giving it a unique position in the Asian AI hardware ecosystem.
South Korea's national push for AI sovereignty, combined with global demand for non-Nvidia AI inference solutions, has created strong tailwinds for Rebellions' data center and edge NPU products.
Rebellions merged with fellow Korean AI chip startup Sapeon in 2024 and has secured significant funding to accelerate development of its next-generation data center inference chip, backed by major Korean institutional investors.
Ricursive Intelligence is building AI that can automate the entire process of designing and testing computer chips, from initial concept to final verification.
Their unique angle is using AI to design better AI chips, which in turn makes the AI smarter — creating a self-improving cycle. While other AI-for-chip-design startups focus on individual steps in the process, Ricursive targets the full chip design and verification pipeline end-to-end, aiming for a level of automation that could democratize custom silicon creation. The massive demand for custom AI accelerators, combined with the prohibitive cost and time of traditional chip design, creates a perfect market opportunity for AI-automated semiconductor development.
Ricursive Intelligence launched with a $35M seed round led by Sequoia Capital in Q4 2025, with a $300M Series A following quickly in January 2026, valuing the company at $4B.
Salience Labs builds processors that use light to perform the core math operations that power AI — specifically the massive matrix multiplications at the heart of every neural network. By doing these calculations with photons instead of electrons, they aim to deliver dramatic improvements in speed and energy efficiency. Salience Labs' photonic processors are designed to perform the matrix multiplication operations central to neural networks using light, which can execute these calculations in a single pass rather than through multiple clock cycles. This approach promises orders of magnitude improvement in energy efficiency for specific AI operations.
The exponential growth of AI model sizes and inference demand is creating an urgent need for fundamentally more energy-efficient computing approaches, which Salience Labs' photonic technology is designed to deliver.
Salience Labs raised $35M in September 2025.
SambaNova builds complete AI computing systems — custom chips paired with software and pre-trained models — that let large organizations run AI privately within their own infrastructure. Their systems are used by government agencies, banks, and energy companies that need powerful AI but can't send sensitive data to outside cloud providers. SambaNova's Reconfigurable Dataflow Unit (RDU) architecture is purpose-built for how data flows through AI models, unlike general-purpose GPUs, providing superior inference efficiency and the ability to handle massive models on a single system. Their three-tier memory hierarchy allows rapid model switching and enables organizations to run multiple AI workloads without sending data to external clouds.
The explosion of enterprise demand for running large language models privately and efficiently has driven SambaNova to refocus on inference economics, positioning its platform as a cost-effective alternative for agentic AI workloads at scale.
SambaNova secured a $350M Series E led by Vista Equity Partners in early 2026, with Intel investing up to $150M as part of a multi-year partnership, after previously being in advanced acquisition talks with Intel at a $1.6B valuation.
SiMa.ai designs computer chips that let robots, vehicles, factory equipment, and other physical devices run AI software directly on the device — without needing to send data to the cloud. Their chips are built to be extremely power-efficient, using a fraction of the energy that traditional processors require. While competitors like Nvidia focus on high-power data center GPUs, SiMa.ai targets the 5W-25W embedded edge segment with a software-centric architecture that delivers 10x better performance per watt. Their platform runs entire ML application pipelines on a single chip, eliminating the host CPU bottleneck that plagues rival solutions.
The rapid expansion of generative AI to edge devices — enabling robots, drones, and vehicles to run LLMs, vision language models, and multimodal AI locally — is driving surging demand for SiMa.ai's low-power silicon.
SiMa.ai raised an $85M funding round led by Maverick Capital in August 2025, bringing total funding to $355M, and began shipping its second-gen Modalix MLSoC platform along with the LLiMa on-device framework for running large language models at the edge.
Snowcap Compute is building AI chips using superconductors, which are materials that allow current to flow without electrical resistance. For superconducting chips to work, the chips need to be kept very cold in cryogenic coolers, which have been inhibitive for cost and power in the past. The demand for AI compute means that the performance unlock of superconducting chips now outweights the operational complexity of cryogenic coolers.
The rapid deployment of AI GPU clusters consuming 40-100+ kilowatts per rack is overwhelming traditional data center cooling infrastructure, creating an urgent market for Snowcap's AI-optimized thermal solutions.
Snowcap Compute raised a $23M Seed round in June 2025, led by Playground Global, with support from Cambium Capital, Page One and others.
Standard Kernel uses AI to autonomously generate highly specialized GPU kernels — the foundational units of computation that determine how efficiently AI models run on hardware. By optimizing down to native chip instructions, it replaces static one-size-fits-all libraries with code tailored to specific workloads and hardware configurations, without requiring changes to models or hardware. While kernel generation has become a popular LLM benchmark, most approaches target higher-level abstractions or simple workloads — Standard Kernel operates at the instruction level to match or beat human-engineered implementations. In partner testing, it demonstrated 80% to 4x performance improvements on H100 GPUs, outperforming Nvidia's cuDNN library in certain scenarios.
As AI models grow larger and inference costs dominate budgets, the efficiency of GPU kernel code has become a critical determinant of AI economics, making Standard Kernel's optimization expertise increasingly valuable.
Standard Kernel raised a $20M seed led by Jump Capital, with participation from General Catalyst, Felicis, CoreWeave, and Ericsson Ventures, along with notable angels including Jeff Dean and SemiAnalysis founder Dylan Patel. The company is early-stage, founded by a team out of MIT and Stanford that created the widely used KernelBench and Kernel Tree Search open-source benchmarks.

Starcloud is building data centers in space, using satellites equipped with solar panels, radiation shielding, and radiative cooling systems to run GPU compute in orbit. Starcloud launched its first satellite, Starcloud-1, in November 2025, carrying an Nvidia H100 GPU. Their next launch, Starcloud 2, is coming later this year. The satellite will contain GPUs, an AWS server, and a bitcoin mining computer. Starcloud raised a $200M total across a $30M Seed and a $170M Series A in March 2026, the latter led by Benchmark and EQT Ventures, vaulting the company to a $1.1B valuation.
Substrate is developing a new chipmaking tool that uses X-ray light to etch the microscopic patterns needed to produce advanced computer chips to replace existing lithography machines. The startup's bigger ambition is to build its own chip factories in the United States and slash the cost of making the world's most advanced processors.
Substrate's approach uses X-ray light from a compact particle accelerator to achieve comparable resolution at dramatically lower cost than its core competitor, ASML. The company plans to not just sell equipment but build its own network of semiconductor fabs, vertically integrating the manufacturing chain. AI's insatiable demand for advanced chips has made domestic semiconductor manufacturing a national security imperative, directly fueling investor and government interest in Substrate's mission to break the ASML/TSMC duopoly.
Substrate emerged from stealth in October 2025 with $100M in funding led by Peter Thiel's Founders Fund at a $1B valuation, with backing from In-Q-Tel and General Catalyst.
TensorWave operates a cloud computing service built entirely on AMD hardware. They specialize in optimizing the AMD software stack so customers can run AI workloads on AMD hardware without the compatibility headaches of doing it themselves. While nearly all GPU clouds are built on Nvidia hardware, TensorWave is one of the few providers betting heavily on AMD's AI accelerators, offering customers an alternative that can be more cost-effective for certain workloads. Their deep optimization of the AMD software stack gives customers production-ready AMD GPU access without the ecosystem challenges of DIY deployment.
The AI industry's desire to reduce dependency on Nvidia as the sole GPU supplier has created a growing market for AMD-based alternatives, which TensorWave serves with purpose-optimized cloud infrastructure.
TensorWave raised nearly $150M of Series A venture funding in a deal led by AMD Ventures, Nexus Venture Partners, and Magnetar Capital. They have deployed one of the largest AMD MI300X GPU clusters for commercial cloud use, establishing itself as a pioneer in the AMD AI cloud ecosystem.

The Biological Computing Company (TBC) is a San Francisco-based startup, founded in 2022 by neurosurgeon-neuroscientists Alex Ksendzovsky (CEO) and Jon Pomeraniec (COO), the company is building a computing platform that positions biology as a complementary compute substrate — not a replacement for silicon, but a new layer designed to improve the efficiency, stability, and adaptability of existing AI systems.
TBC's core platform encodes real-world data — images, video, and other signals — into living neuronal cultures grown on electrode grids. The neural responses are then decoded into machine-readable representations that are mapped onto frontier AI models through modular adapters. In parallel, the company's Algorithm Discovery platform observes the computational patterns of living neural networks to extract learning principles, using those insights to inform the design of AI architectures beyond the transformer paradigm.
The company reported a 23x retained improvement in video model efficiency using its biological adapters. TBC targets compute-intensive domains including computer vision, generative video, and dynamic world models — application areas where conventional scaling has encountered economic and energy constraints. The underlying thesis is that biological neural networks, refined through billions of years of evolution, offer efficiency and generalization properties that silicon-based systems do not inherently replicate.
TBC emerged from stealth in February 2026 with a $25M seed round led by Primary Venture Partners, with participation from Builders VC, E1 Ventures, Proximity, Refactor Capital, Tusk Ventures, and Wonder Ventures. The company has opened a flagship lab in San Francisco's Mission Bay.
* Primary is an investor in The Biological Computing Company
Together AI operates a cloud platform focused on open-source AI models, letting companies train, customize, and run models like Llama and Mixtral without building their own infrastructure. They also invest heavily in AI research, contributing to major open-source projects that benefit the broader community. Together AI has positioned itself as the leading champion of the open-source AI ecosystem, investing heavily in research and contributing to major open models while building the most optimized infrastructure for running them. Their research team has co-authored influential papers and developed key training and inference optimizations that benefit the entire open-source AI community.
The explosive adoption of open-source AI models like Llama, Mixtral, and DeepSeek across enterprises has driven Together AI's growth as the go-to platform for open-source model deployment and customization.
Together AI raised a $305M Series B in 2025 at a $3.3B valuation led by Salesforce Ventures and Nvidia, and has continued to expand its research contributions and enterprise partnerships.
Unconventional AI is designing a new kind of computer chip inspired by how the human brain works, using analog circuits instead of the traditional digital ones found in every computer today.
The idea is that by letting the physics of silicon do the computing directly — rather than forcing everything into ones and zeros — they can make AI run on a tiny fraction of the energy it uses now. Rather than engineering circuits to behave digitally (ones and zeros), Unconventional AI exploits the natural non-linear dynamics of analog silicon to perform computation, potentially achieving 1,000x better efficiency than current chips. This approach builds intelligence directly on the physics of the substrate, mirroring how biological neurons operate on just 20 watts.
Unconventional AI launched in December 2025 with a staggering $475M seed round at a $4.5B valuation, led by Lightspeed and Andreessen Horowitz with participation from Jeff Bezos — one of the largest seed rounds in venture history.
Upscale AI builds open-standard networking infrastructure — silicon, systems, and software — purpose-built for AI workloads like GPU-to-GPU communication at ultra-low latency. Its SkyHammer architecture unifies GPUs, accelerators, memory, and storage into a single fabric, built on SONiC, Ultra Ethernet, and UALink as an open alternative to Nvidia's NVLink. Upscale is fully vertically integrated across the networking stack. It also bets entirely on open standards rather than proprietary protocols, aiming to give hyperscalers and enterprises vendor-neutral interoperability instead of lock-in.
With GPU compute remaining expensive and scarce, organizations are increasingly motivated to maximize the utilization and performance of their existing AI infrastructure, creating strong demand for Upscale AI's optimization solutions.
Upscale has raised over $300M across a $100M+ seed and $200M Series A from Tiger Global, Premji Invest, Intel Capital, Qualcomm Ventures, and others. Its products are targeting hyperscalers and AI infrastructure operators, with networking solutions slated to begin shipping in 2026.

VAST Data builds the data foundation that large AI clusters run on. When tens of thousands of GPUs train a model, the bottleneck is often not the chips but getting data to them fast enough; VAST's platform is designed to keep that pipeline saturated, and has since expanded upward into databases, search and agent infrastructure, what the company markets as an AI Operating System.
Unlike conventional storage architectures that force a tradeoff between performance and capacity, VAST's disaggregated shared-everything design separates compute from storage enclosures over NVMe fabric, letting a single namespace behave like fast flash for hot data while scaling to exabytes. Its closest venture comp, WEKA, attacks the same GPU-starvation problem; VAST's differentiation is the move beyond storage into the layers above it, selling database, InsightEngine and AgentEngine components into accounts already anchored on the platform.
The AI datacenter buildout has made the data layer strategically valuable, and VAST has captured an unusual share of the largest deployments. It underpins xAI's Colossus cluster, and CoreWeave signed a $1.17B multi-year agreement in November 2025 making VAST the primary data foundation for its Nvidia-accelerated cloud. The company reports surpassing $4B in cumulative bookings and exiting its last fiscal year above $500M in committed ARR with positive free cash flow. VAST closed a $1B Series F in April 2026 at a $30B valuation, led by Drive Capital with Access Industries co-leading and Nvidia, Fidelity and NEA participating, more than triple its $9.1B Series E mark from late 2023. Since then VAST has added Hyundai Motor Company and Kia as strategic investors in October 2026 after years of production use at their 42dot autonomy unit, expanded its DigitalOcean partnership to run the VAST AI OS inside DigitalOcean's AI-Native Cloud, and launched an agent-sandbox foundation stack with Nvidia at GTC Berlin.
WEKA is a software-defined data platform company that delivers cloud-native, high-performance storage purpose-built for AI and accelerated compute workloads. Its NeuralMesh architecture interconnects data, compute, and AI services across on-prem, cloud, and edge environments, replacing legacy storage with dynamic data pipelines that keep GPUs continuously fed.
WEKA's architecture eliminates the traditional tradeoff between performance and capacity by providing a single namespace that performs like NVMe flash for hot data and scales like cloud storage for large datasets. Their purpose-built data pipeline can feed thousands of GPUs simultaneously without the I/O bottlenecks that plague conventional storage systems.
AI training requires feeding massive datasets to GPU clusters at speeds that traditional storage cannot sustain; WEKA's high-performance data platform solves this by delivering hundreds of gigabytes per second to AI workloads.
WEKA raised $140M in a Series E at a $1.6B valuation, bringing total funding to ~$375M — notably raised entirely from existing investors. The company has over 300 customers including 12 of the Fortune 50 and AI companies like Stability AI, Midjourney, and ElevenLabs, with ARR exceeding $100M and doubling year-over-year.

xLight is building extremely powerful lasers called free electron lasers that produce the specific type of light needed to manufacture the most advanced computer chips.
Today's chipmaking light sources are running up against physical limits, and xLight's technology could provide four times more light while using less energy and fewer consumable materials. Current EUV light sources from ASML use energy-intensive tin plasma lasers that provide only 25% of the light future lithography demands; xLight's FEL approach directly emits EUV light with programmable wavelengths, eliminating consumables like tin and hydrogen. A single xLight source can feed multiple lithography machines, fundamentally changing fab economics.
The surging demand for AI chips is pushing semiconductor manufacturers to their limits, creating urgent need for more productive and cost-efficient lithography solutions that xLight's technology directly addresses.
xLight closed an oversubscribed $40M Series B led by Playground Global in July 2025, and in December 2025 signed a $150M Letter of Intent with the U.S. Department of Commerce under the CHIPS Act — the first award from the Trump Administration's CHIPS R&D Office.
Akash Systems uses diamond-based cooling technology that eliminates thermal constraints, unlocking higher performance, greater efficiency, longer system life, and lower costs. By bonding synthetic diamond to chip surfaces, they pull heat away far more effectively than traditional methods, allowing devices to run faster and last longer.
Akash Systems' Diamond Cooling fuses the superior thermal conductivity of synthetic diamond materials with the semiconductors that power modern electronics, creating a new material that dissipates heat efficiently, boosts energy efficiency, enhances system-wide performance, and increases the lifespan of electronics.
The massive heat generated by AI training and inference workloads on GPUs has created urgent demand for Akash's Diamond Cooling server technology, which enables 25% overclocking and doubles server lifetimes by eliminating thermal throttling.
Ayar Labs replaces the traditional copper wire connections between chips with tiny light-based links that can move data much faster while using less power. Their optical connectors fit inside standard chip packages, making it possible to dramatically speed up data movement without overhauling existing manufacturing processes.
While other optical interconnect companies focus on longer-distance fiber optic links, Ayar Labs brings optics inside the chip package itself, addressing the data movement bottleneck at its most fundamental level. Their technology is compatible with existing semiconductor manufacturing processes, enabling adoption without requiring entirely new fab infrastructure.
AI training clusters require moving enormous amounts of data between GPUs, CPUs, and memory at speeds that electrical interconnects increasingly cannot deliver, making Ayar Labs' optical I/O a critical enabling technology for next-generation AI systems. Ayar Labs has deepened partnerships with major chipmakers and received investment from strategic players including Nvidia, as demand for its optical I/O technology accelerates alongside AI infrastructure buildouts.
Atomic Machines is building a new type of robotic manufacturing system that can create tiny mechanical devices with features just a few millionths of a meter in size. Their long-term goal is to make manufacturing so precise that objects can be assembled at the atomic level, starting with components for the AI data center market.
Unlike conventional MEMS fabrication tied to silicon fab limitations, Atomic Machines' robotic platform enables rapid prototyping and large-scale production of complex 3D microdevices with single-digit-micron features. Their approach breaks traditional manufacturing constraints, allowing novel device geometries and materials combinations impossible with existing methods.
Atomic Machines has developed its first device targeting the AI data center market, capitalizing on the need for novel micro-components that can improve the performance and efficiency of AI infrastructure. Atomic Machines has raised $144M in total funding backed by investors including Sozo Ventures and XTX Ventures, and received a major California Competes tax credit as part of a $156.3M investment to expand its MEMS fabrication facilities in Santa Clara and Emeryville.
Baseten provides a cloud platform that makes it easy for companies to deploy and run AI models in production without managing the underlying server infrastructure. Their system automatically handles things like assigning the right number of graphics processors and scaling up or down based on demand.
Unlike general-purpose cloud providers, Baseten is built specifically for ML model deployment, offering optimized model serving with features like automatic batching, streaming responses, and built-in GPU scheduling. Their developer-first approach provides simple APIs and Truss, an open-source model packaging framework, that makes deployment drastically simpler than DIY alternatives.
The explosion of companies deploying AI models — from LLMs to image generators to recommendation systems — has created massive demand for inference infrastructure, fueling Baseten's rapid growth as a specialized platform. Baseten reached a $5B valuation with a $300M funding round led by IVP and CapitalG in January 2026.
ChipAgents builds AI agents that help engineers design and verify computer chips dramatically faster — think of it as a team of AI coworkers that can read chip specifications, write and debug design code, and run verification tests autonomously. Their platform plugs into engineers' existing code editors and workflows, aiming to make chip design 10 times more productive.
Unlike traditional chip design tools that assist with individual tasks, ChipAgents deploys coordinated teams of AI agents that take ownership of entire workflows — reading specs, breaking down objectives, implementing solutions, validating results, and iterating without constant human guidance. Their agents have demonstrated faster test generation and faster specification comprehension in production environments at major semiconductor companies.
The breakthroughs in agentic AI — where AI systems can autonomously plan, reason, and execute multi-step tasks — have made it possible for the first time to automate the complex, judgment-heavy work of chip design and verification. ChipAgents raised $74M in total funding including an oversubscribed $50M Series A1 in February 2026 led by TSMC-backed Matter Venture Partners, with Bessemer, Micron, MediaTek, and Ericsson also investing. The company reported 140x year-over-year revenue growth, deployments at 80 semiconductor companies, and advisory board additions including former CEOs of Cadence, Mentor Graphics (Siemens), and the former CTO of Synopsys.
Cognichip is building an AI system that can help engineers design computer chips faster and cheaper. Their goal is to cut chip development time in half and reduce costs by 75%, making it possible for more companies to afford designing their own custom chips.
While existing EDA tools assist with specific design tasks, Cognichip's ACI model acts like an expert chip engineer, understanding and solving design problems with designer-level cognitive abilities across the entire development process. The platform promises faster design cycles and cost reduction through concurrent local and global optimization, unlike the serial approach of traditional tools.
The generative AI revolution has made it possible for the first time to build foundational models that understand semiconductor physics deeply enough to automate and accelerate chip design at scale. Cognichip emerged from stealth in May 2025 with $33M in seed funding co-led by Lux Capital and Mayfield, assembling a team of AI experts from Stanford, Google, MIT, and semiconductor veterans from Amazon, Apple, and Synopsys.
Corintis builds tiny liquid cooling channels that sit directly on or near a chip's surface, pulling heat away right at the source rather than relying on fans or larger cooling systems further away. This approach is becoming essential as AI chips push past power levels that traditional cooling methods can no longer handle.
Traditional cooling solutions work at the system level, but Corintis' microfluidic technology operates at the chip level, channeling coolant through microscale structures directly on or near the silicon die. This removes heat much closer to its source than conventional approaches, enabling chips to run at higher power levels without thermal throttling.
Corintis raised $24M in Series A funding in late 2025, adding Intel CEO Lip Bu Tan to their board. They partnered with Microsoft to show 3x better cooling with microfluidics versus cooling alternatives.
Cornelis Networks makes high-speed networking equipment designed specifically for connecting thousands of AI processors together in a computing cluster. Their technology is optimized for the unique communication patterns of AI training, where processors constantly need to share data with each other at extremely low latency.
Cornelis' networking fabric is purpose-built for the collective communication patterns of AI training workloads, where thousands of GPUs must synchronize gradients and share data in patterns very different from traditional data center traffic. Their architecture delivers lower latency and higher message rate than general-purpose Ethernet for AI-specific communication. The company formed as a spinout of Intel's OmniPath networking product.
Crusoe builds AI data centers near renewable and stranded energy sources, reducing both cost and carbon footprint compared to conventional cloud providers. Their first product to market used stranded energy, like natural gas being burned off at oil wells or renewable energy that the grid can't absorb, to power colocated data centers.
The company operates H100 and H200 GPU clusters and is a preferred provider for major AI workloads, including a 2.1GW data center in Abilene, TX to support Microsoft.
The enormous and growing power demands of AI training and inference have made cheap energy access the key competitive advantage in AI infrastructure, perfectly aligning with Crusoe's stranded energy model. Crusoe raised a $1.3B Series E, led by Valor Equity Partners and Mubadala Capital in October 2025.
* Primary is an investor in Crusoe, through the acquisition of Atero.
d-Matrix builds processors that perform AI calculations directly inside the memory where data is stored, rather than constantly shuttling data back and forth between separate memory and processing units. This approach dramatically reduces the energy wasted on moving data around, which is one of the biggest inefficiencies in running AI today.
Unlike GPUs that shuttle data between separate compute and memory units, d-Matrix performs calculations directly inside memory arrays using a digital (not analog) approach, maintaining precision while dramatically reducing data movement energy. This gives them both the efficiency benefits of in-memory computing and the reliability of digital circuits.
As AI inference workloads grow to dominate data center spending — projected to surpass training costs — d-Matrix's inference-optimized architecture offers compelling total cost of ownership advantages over GPU-based solutions. d-Matrix raised a $275M Series C in Q4 2025 led by BullhoundCapital, Triatomic Capital, and Temasek, with participation from Microsoft's M12 and the Qatar Investment Authority.
Eliyan makes high-speed connection technology that lets multiple small chips inside a processor package communicate with each other as if they were a single large chip. Their approach works with standard, affordable packaging materials rather than requiring expensive specialty components, making chiplet-based designs more practical.
Unlike silicon interposers or bridges that are expensive and supply-constrained, Eliyan's interconnect works on standard organic packaging substrates while still achieving the bandwidth density typically associated with advanced packaging. This means chiplet-based designs can achieve high performance without the cost premium of exotic packaging technologies.
The shift to chiplet-based AI processor designs — combining specialized compute, memory, and networking dies — is creating strong demand for Eliyan's high-bandwidth interconnect technology that makes chiplet integration practical and affordable. Eliyan raised $62M in Series B funding in 2025 and announced key partnerships with semiconductor manufacturers, as its NuLink technology gained traction as a critical enabler of next-generation chiplet architectures.
Eridu is building an AI networking solution, designed from the ground up for scale-out and scale-up networks in AI data centers. Eridu’s network switch delivers the order-of-magnitude advance in performance, radix and throughput needed to maximize utilization, throughput and efficiency in AI clusters and data centers.
Demands for data movement are growing exponentially as hyperscalers, frontier labs and neoclouds build ever-larger AI data centers, exposing fragility in the networking stack that exists today. At the same time, new architectural and algorithmic innovations in AI are only making the network bottleneck worse.
Eridu raised an oversubscribed $200M Series A in March 2026, led by Socratic Partners, John Doerr, Hudson River Trading, Capricorn Investment Group and Matter Venture Partners, with participation by Bosch Ventures, Eclipse Capital, Fusion Fund, Osage University Partners, SBVA, TDK Ventures, VentureTech Alliance, and Zelda Ventures among others.
Etched is a Cupertino-based semiconductor startup, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen. The company is building frontier inference clusters. Their rack-scale systems leverage two core architectures, Low Voltage Inference and Cluster Scale Memory, so frontier models can run performantly across both prefill and decode workloads.
Etched has raised $800M in funding, with investors including Stripes, Ribbit Capital, Positive Sum, and Primary Venture Partners.
* Primary is an investor in Etched
Fireworks AI runs a platform where developers can access, customize, and deploy open-source AI models through a fast and affordable cloud service. They have deeply optimized how models run on graphics processors, squeezing significantly more speed out of each chip than standard approaches.
Fireworks differentiates through its compound AI system approach, allowing developers to compose multiple models and tools together in a single API call, and through deep kernel-level optimizations that squeeze more performance from each GPU. Their FireAttention engine and custom CUDA kernels deliver significantly faster token generation than standard serving frameworks.
The rapid proliferation of open-source AI models and enterprise adoption of LLMs has created a massive market for optimized inference platforms, driving Fireworks' growth as developers seek faster, cheaper alternatives to self-hosting. Fireworks AI raised $250M at a $4B valuation in October 2025 in a Series C led by Lightspeed Venture Partners, Index Ventures and Evantic.
Firmus Technologies builds and operates data centers designed from the ground up to handle AI workloads, with advanced power and cooling systems optimized for high-performance computing. Their facilities are purpose-built for AI rather than being traditional data centers retrofitted with graphics processors.
Firmus emphasizes operational efficiency and energy sustainability in its data center design, incorporating advanced power and cooling technologies to minimize environmental impact while maximizing compute density. Their facilities are designed from the ground up for AI workloads rather than being retrofitted from general-purpose data centers.
The surge in demand for AI training and inference capacity has created a massive need for purpose-built AI data centers, positioning Firmus as part of the new wave of AI-native infrastructure providers. Firmus raised nearly $550M in capital from Nvidia and others in September 2025. Soon after, they raised $10B in debt in January 2026 to fund Project Southgate, aimed to deploy infrastructure for sovereign AI in Australia.
Fractile is a UK-based startup designing chips that run large language models more efficiently by rethinking how numbers are represented and processed during AI calculations. Their approach squeezes more useful computation out of every watt of power, targeting the growing cost of running AI at scale. Fractile's architecture exploits the mathematical properties of LLMs to represent computations more efficiently than conventional floating-point approaches, extracting more useful work per watt than GPU-based solutions. Their approach targets the specific mathematical operations that dominate transformer inference, rather than trying to be a general-purpose accelerator.
The rapid scaling of LLM deployment across enterprise and consumer applications is creating massive demand for more efficient inference hardware, which Fractile's numerically optimized architecture directly addresses.
Fractile announced it will be investing over $100M into building out semiconductor capacity in the UK.
FluidStack runs a distributed cloud that pulls together GPUs from data centers around the world, giving AI developers on-demand access to computing power at competitive prices. Their marketplace model aggregates capacity from multiple partners, often making hard-to-find AI chips more available than on the big cloud platforms. Unlike traditional cloud providers with centralized data centers, FluidStack's distributed model aggregates GPU capacity from multiple partners, enabling more competitive pricing and greater availability of hard-to-find AI accelerators. Their marketplace approach creates price competition among GPU suppliers, driving costs down for end users.
The chronic shortage of GPU compute for AI training has driven developers and companies to seek alternative cloud providers like FluidStack that can aggregate distributed capacity and offer better availability than hyperscalers.
FluidStack raised $700M in December 2025 from Situational Awareness, valuing the company at $7B. They are reportedly in talks to raise at an $18B valuation, as of April 2026.
Gimlet Labs is building what it calls the first "multi-silicon inference cloud" — software that allows AI workloads to run simultaneously across diverse hardware types including CPUs, GPUs, and specialized accelerators. Its platform automatically disaggregates agentic workloads into component stages and maps each to the most suitable chip without requiring code changes. Rather than optimizing for a single chip vendor, Gimlet treats the entire heterogeneous data center fleet as one unified compute pool — addressing the problem that existing hardware runs at only 15 to 30 percent utilization. It partners across Nvidia, AMD, Intel, ARM, Cerebras, and d-Matrix, positioning itself as the VMware of AI inference rather than a chip-specific stack.
As AI moves from experimentation to production deployment across enterprises, the need for robust operational infrastructure and tooling is growing rapidly, directly driving demand for Gimlet Labs' solutions.
Gimlet raised an $80M Series A led by Menlo Ventures, bringing total funding to $92M, with angels including Intel CEO Lip-Bu Tan and former VMware CEO Raghu Raghuram. The company launched publicly in October 2025 with eight-figure revenues out of the gate and has since more than doubled its customer base across frontier AI labs and major cloud providers.
Lambda provides cloud computing and hardware specifically built for AI researchers and developers, offering ready-to-use servers loaded with the latest graphics processors and pre-configured AI software. They have become a go-to choice for machine learning teams that want powerful computing without the setup complexity of traditional cloud providers. Lambda is laser-focused on the ML/AI community rather than being a general cloud provider, offering pre-configured software stacks and developer-friendly tooling that eliminate the setup friction researchers face on traditional clouds. Their 1-Click Clusters feature enables teams to spin up multi-node GPU training infrastructure in minutes rather than weeks.
The explosive growth of AI research and commercial model development has driven surging demand for Lambda's ML-optimized cloud and hardware products, establishing it as the go-to GPU cloud for AI practitioners.
In November 2025, they raised $1.6B from TWG Global in a Series E funding round. As of January 2026, Lambda Labs was in talks to raise $350M in pre-IPO funding.
Lightmatter builds networking equipment that uses light instead of electricity to connect AI chips across a data center, delivering dramatically more bandwidth with less energy. Their technology can link chips together at speeds previously only possible within a single processor, removing the networking bottleneck that slows down large AI systems. Lightmatter's Passage interconnect is capable of connecting chips across an entire data center with the bandwidth density previously only possible within a single chip. Their programmable photonic technology can dynamically reconfigure data pathways, unlike fixed electrical connections.
As AI clusters scale to hundreds of thousands of GPUs, the network connecting them becomes the primary bottleneck; Lightmatter's photonic fabric directly solves this by providing orders of magnitude more bandwidth at lower power.
Lightmatter raised $400M in 2025, reaching a $4.4B valuation, and has been deepening partnerships with major AI chip and cloud companies to integrate its photonic interconnect technology into next-generation data center architectures.
Majestic Labs builds AI servers with custom accelerator and memory interface chips that disaggregate memory from compute, enabling up to 128 TB of high-bandwidth memory per server. Each server is designed to collapse multiple racks of conventional GPU equipment into a single unit for training and inference workloads. Majestic attacks the "memory wall" directly by delivering roughly 1,000x the memory capacity of a standard GPU server. The architecture is fully vertically integrated (custom silicon, systems, software) and maintains full programmability to support workloads beyond current transformer models.
The insatiable demand for more efficient AI computing has created opportunities for companies like Majestic Labs that bring fresh architectural thinking to the fundamental challenges of AI hardware design.
Majestic has raised over $100M in Series A financing and was founded by the team behind Meta's FAST silicon group and Google's GChips, holding 120+ patents collectively. The company is targeting hyperscalers and enterprises in data-intensive verticals like financial services and pharma.
MatX is a San Francisco-based AI chip company, founded in 2023, designing high-throughput accelerators specifically optimized for running large language models. The company went from a $25M seed round to over $600M in total funding in under two years. Their first chip, MatX One, is based on a splittable systolic array which combines the low latency of SRAM-first designs with the long-context support of HBM.
The MatX founding team comes from Google's TPU team. MatX raised a $500M round in 2026 and is building chips designed to deliver more compute per dollar for LLM inference than general-purpose GPUs.
Neurophos is developing chips that combine light-based computing elements with traditional electronic circuits on the same piece of silicon. This hybrid approach aims to deliver the speed and energy advantages of computing with light while remaining compatible with existing chip manufacturing infrastructure. Neurophos' approach integrates photonic computing elements with conventional electronic circuits on the same chip, offering a practical path to adopting optical computing without requiring a complete overhaul of existing infrastructure. Their hybrid approach balances the speed of light-based processing with the flexibility of electronics.
The energy and performance demands of AI inference and training are increasingly exceeding what purely electronic chips can deliver, creating a compelling use case for Neurophos' photonic computing technology.
In January 2026, Neurophos raised a $110M Series A, led by Gates Frontier, with participation from M12, Carbon Direct Capital, Silicon Catalyst Ventures, and more.
Modal is a cloud platform that lets developers run AI and data-intensive code on powerful remote servers using nothing more than a few lines of Python — no infrastructure setup required. The platform instantly provisions the right hardware, including graphics processors, and scales automatically. Modal's key innovation is making cloud GPU access feel like running code locally — developers write normal Python functions and Modal handles everything from containerization to GPU allocation in milliseconds. Unlike traditional cloud providers that require provisioning VMs and managing Kubernetes, Modal's serverless approach means zero infrastructure management.
The AI boom has created millions of developers who need GPU access for model training, fine-tuning, and inference but lack infrastructure expertise, making Modal's frictionless cloud platform increasingly essential.
Modal announced a $366M Series C in May 2026, reaching a $4.65B valuation, as its sandboxes product has taken off, pushing the business past $300M ARR.
Nexthop AI builds custom networking equipment — switches, software, and optical connections — designed specifically for the massive AI data centers operated by the world's largest cloud companies. They have a unique co-development model, focused on delivering hardware and software purpose-built cloud providers running open source operating systems. Unlike traditional networking vendors that sell one-size-fits-all products, Nexthop co-engineers custom solutions alongside each hyperscaler's own team, integrating seamlessly into their specific technology stack. CEO Anshul Sadana previously spent 17 years at Arista Networks, serving as its Chief Customer Officer, then Chief Operating Officer.
AI training clusters require thousands of processors to exchange data at blinding speeds, making the network fabric between them a critical performance bottleneck — and Nexthop's purpose-built AI networking equipment directly addresses this growing need.
Nexthop AI raised an oversubscribed $500M Series B in March 2026 led by Lightspeed Venture Partners with Andreessen Horowitz joining, catapulting the company's valuation to $4.2 billion — just one year after emerging from stealth with $110M in funding. The company simultaneously unveiled three new Ethernet switch platforms engineered for hyperscale AI clusters and neoclouds.
PowerLattice makes a small chip component called a power delivery chiplet that sits inside a processor's package and delivers electricity right next to where the computing happens. By moving power regulation closer to the chip, they cut energy waste by more than half — a big deal when AI chips are consuming record amounts of electricity. While competitors deliver power from the motherboard inches away from the chip, PowerLattice places voltage regulation just hundreds of micrometers from the processor die, cutting compute power needs by more than 50%. Their chiplet-based approach is inherently scalable and can be deployed in parallel, adapting to any SoC power topology.
AI accelerators are straining data center power and cooling infrastructure, making PowerLattice's 50%+ power reduction directly critical to sustaining AI compute scaling.
PowerLattice emerged from stealth in November 2025 with a $25M Series A led by Playground Global and Celesta Capital (with former Intel CEO Pat Gelsinger as a board member).
Nscale is a European cloud provider building AI computing infrastructure across Europe, giving companies and governments access to powerful graphics processors while keeping their data on European soil. This addresses growing demand from organizations that need AI computing power but must comply with European data privacy regulations. Nscale positions itself as Europe's answer to U.S.-dominated GPU cloud providers, offering data sovereignty and compliance with EU regulations like GDPR for organizations that cannot or prefer not to send data to American hyperscalers. Their European-first approach addresses a critical gap in the market for sovereign AI infrastructure.
European companies and governments increasingly need local, sovereign AI compute infrastructure to comply with data regulations and reduce dependence on U.S. cloud providers, directly driving Nscale's growth.
In March 2026, Nscale and Microsoft announced a collaboration to to deliver 1.35GW of Vera Rubin NVIDIA capacity on a new West Virginia Campus.
PsiQuantum is building a large-scale quantum computer that uses particles of light instead of the electrical circuits used in traditional computers. Their key advantage is that they manufacture their quantum components in standard chip factories, making their approach uniquely scalable compared to other quantum products. Unlike competitors using superconducting or trapped-ion qubits that require exotic manufacturing, PsiQuantum fabricates its quantum components in standard silicon foundries, making its approach uniquely scalable to the millions of qubits needed for useful computation. Their photonic architecture operates at room temperature components where possible, avoiding the extreme cooling requirements of superconducting approaches.
Quantum computing promises to solve optimization and simulation problems that are intractable for classical AI, and PsiQuantum's photonic approach could enable quantum-enhanced machine learning algorithms that dramatically accelerate AI capabilities.
PsiQuantum raised $1.5B in a Series E at a $10.5B valuation, with investors including Nvidia, Blackrock, and Baillie Gifford.
Rebellions is a South Korean chipmaker designing processors built specifically for running AI workloads in both data centers and smaller devices at the network edge. As the leading AI chip company in South Korea, Rebellions benefits from strong government backing and partnerships with Samsung Foundry for manufacturing, giving it a unique position in the Asian AI hardware ecosystem.
South Korea's national push for AI sovereignty, combined with global demand for non-Nvidia AI inference solutions, has created strong tailwinds for Rebellions' data center and edge NPU products.
Rebellions merged with fellow Korean AI chip startup Sapeon in 2024 and has secured significant funding to accelerate development of its next-generation data center inference chip, backed by major Korean institutional investors.
Ricursive Intelligence is building AI that can automate the entire process of designing and testing computer chips, from initial concept to final verification.
Their unique angle is using AI to design better AI chips, which in turn makes the AI smarter — creating a self-improving cycle. While other AI-for-chip-design startups focus on individual steps in the process, Ricursive targets the full chip design and verification pipeline end-to-end, aiming for a level of automation that could democratize custom silicon creation. The massive demand for custom AI accelerators, combined with the prohibitive cost and time of traditional chip design, creates a perfect market opportunity for AI-automated semiconductor development.
Ricursive Intelligence launched with a $35M seed round led by Sequoia Capital in Q4 2025, with a $300M Series A following quickly in January 2026, valuing the company at $4B.
SambaNova builds complete AI computing systems — custom chips paired with software and pre-trained models — that let large organizations run AI privately within their own infrastructure. Their systems are used by government agencies, banks, and energy companies that need powerful AI but can't send sensitive data to outside cloud providers. SambaNova's Reconfigurable Dataflow Unit (RDU) architecture is purpose-built for how data flows through AI models, unlike general-purpose GPUs, providing superior inference efficiency and the ability to handle massive models on a single system. Their three-tier memory hierarchy allows rapid model switching and enables organizations to run multiple AI workloads without sending data to external clouds.
The explosion of enterprise demand for running large language models privately and efficiently has driven SambaNova to refocus on inference economics, positioning its platform as a cost-effective alternative for agentic AI workloads at scale.
SambaNova secured a $350M Series E led by Vista Equity Partners in early 2026, with Intel investing up to $150M as part of a multi-year partnership, after previously being in advanced acquisition talks with Intel at a $1.6B valuation.
Salience Labs builds processors that use light to perform the core math operations that power AI — specifically the massive matrix multiplications at the heart of every neural network. By doing these calculations with photons instead of electrons, they aim to deliver dramatic improvements in speed and energy efficiency. Salience Labs' photonic processors are designed to perform the matrix multiplication operations central to neural networks using light, which can execute these calculations in a single pass rather than through multiple clock cycles. This approach promises orders of magnitude improvement in energy efficiency for specific AI operations.
The exponential growth of AI model sizes and inference demand is creating an urgent need for fundamentally more energy-efficient computing approaches, which Salience Labs' photonic technology is designed to deliver.
Salience Labs raised $35M in September 2025.
SiMa.ai designs computer chips that let robots, vehicles, factory equipment, and other physical devices run AI software directly on the device — without needing to send data to the cloud. Their chips are built to be extremely power-efficient, using a fraction of the energy that traditional processors require. While competitors like Nvidia focus on high-power data center GPUs, SiMa.ai targets the 5W-25W embedded edge segment with a software-centric architecture that delivers 10x better performance per watt. Their platform runs entire ML application pipelines on a single chip, eliminating the host CPU bottleneck that plagues rival solutions.
The rapid expansion of generative AI to edge devices — enabling robots, drones, and vehicles to run LLMs, vision language models, and multimodal AI locally — is driving surging demand for SiMa.ai's low-power silicon.
SiMa.ai raised an $85M funding round led by Maverick Capital in August 2025, bringing total funding to $355M, and began shipping its second-gen Modalix MLSoC platform along with the LLiMa on-device framework for running large language models at the edge.
Snowcap Compute is building AI chips using superconductors, which are materials that allow current to flow without electrical resistance. For superconducting chips to work, the chips need to be kept very cold in cryogenic coolers, which have been inhibitive for cost and power in the past. The demand for AI compute means that the performance unlock of superconducting chips now outweights the operational complexity of cryogenic coolers.
The rapid deployment of AI GPU clusters consuming 40-100+ kilowatts per rack is overwhelming traditional data center cooling infrastructure, creating an urgent market for Snowcap's AI-optimized thermal solutions.
Snowcap Compute raised a $23M Seed round in June 2025, led by Playground Global, with support from Cambium Capital, Page One and others.
Standard Kernel uses AI to autonomously generate highly specialized GPU kernels — the foundational units of computation that determine how efficiently AI models run on hardware. By optimizing down to native chip instructions, it replaces static one-size-fits-all libraries with code tailored to specific workloads and hardware configurations, without requiring changes to models or hardware. While kernel generation has become a popular LLM benchmark, most approaches target higher-level abstractions or simple workloads — Standard Kernel operates at the instruction level to match or beat human-engineered implementations. In partner testing, it demonstrated 80% to 4x performance improvements on H100 GPUs, outperforming Nvidia's cuDNN library in certain scenarios.
As AI models grow larger and inference costs dominate budgets, the efficiency of GPU kernel code has become a critical determinant of AI economics, making Standard Kernel's optimization expertise increasingly valuable.
Standard Kernel raised a $20M seed led by Jump Capital, with participation from General Catalyst, Felicis, CoreWeave, and Ericsson Ventures, along with notable angels including Jeff Dean and SemiAnalysis founder Dylan Patel. The company is early-stage, founded by a team out of MIT and Stanford that created the widely used KernelBench and Kernel Tree Search open-source benchmarks.
TensorWave operates a cloud computing service built entirely on AMD hardware. They specialize in optimizing the AMD software stack so customers can run AI workloads on AMD hardware without the compatibility headaches of doing it themselves. While nearly all GPU clouds are built on Nvidia hardware, TensorWave is one of the few providers betting heavily on AMD's AI accelerators, offering customers an alternative that can be more cost-effective for certain workloads. Their deep optimization of the AMD software stack gives customers production-ready AMD GPU access without the ecosystem challenges of DIY deployment.
The AI industry's desire to reduce dependency on Nvidia as the sole GPU supplier has created a growing market for AMD-based alternatives, which TensorWave serves with purpose-optimized cloud infrastructure.
TensorWave raised nearly $150M of Series A venture funding in a deal led by AMD Ventures, Nexus Venture Partners, and Magnetar Capital. They have deployed one of the largest AMD MI300X GPU clusters for commercial cloud use, establishing itself as a pioneer in the AMD AI cloud ecosystem.
Substrate is developing a new chipmaking tool that uses X-ray light to etch the microscopic patterns needed to produce advanced computer chips to replace existing lithography machines. The startup's bigger ambition is to build its own chip factories in the United States and slash the cost of making the world's most advanced processors.
Substrate's approach uses X-ray light from a compact particle accelerator to achieve comparable resolution at dramatically lower cost than its core competitor, ASML. The company plans to not just sell equipment but build its own network of semiconductor fabs, vertically integrating the manufacturing chain. AI's insatiable demand for advanced chips has made domestic semiconductor manufacturing a national security imperative, directly fueling investor and government interest in Substrate's mission to break the ASML/TSMC duopoly.
Substrate emerged from stealth in October 2025 with $100M in funding led by Peter Thiel's Founders Fund at a $1B valuation, with backing from In-Q-Tel and General Catalyst.
Together AI operates a cloud platform focused on open-source AI models, letting companies train, customize, and run models like Llama and Mixtral without building their own infrastructure. They also invest heavily in AI research, contributing to major open-source projects that benefit the broader community. Together AI has positioned itself as the leading champion of the open-source AI ecosystem, investing heavily in research and contributing to major open models while building the most optimized infrastructure for running them. Their research team has co-authored influential papers and developed key training and inference optimizations that benefit the entire open-source AI community.
The explosive adoption of open-source AI models like Llama, Mixtral, and DeepSeek across enterprises has driven Together AI's growth as the go-to platform for open-source model deployment and customization.
Together AI raised a $305M Series B in 2025 at a $3.3B valuation led by Salesforce Ventures and Nvidia, and has continued to expand its research contributions and enterprise partnerships.
Unconventional AI is designing a new kind of computer chip inspired by how the human brain works, using analog circuits instead of the traditional digital ones found in every computer today.
The idea is that by letting the physics of silicon do the computing directly — rather than forcing everything into ones and zeros — they can make AI run on a tiny fraction of the energy it uses now. Rather than engineering circuits to behave digitally (ones and zeros), Unconventional AI exploits the natural non-linear dynamics of analog silicon to perform computation, potentially achieving 1,000x better efficiency than current chips. This approach builds intelligence directly on the physics of the substrate, mirroring how biological neurons operate on just 20 watts.
Unconventional AI launched in December 2025 with a staggering $475M seed round at a $4.5B valuation, led by Lightspeed and Andreessen Horowitz with participation from Jeff Bezos — one of the largest seed rounds in venture history.
Upscale AI builds open-standard networking infrastructure — silicon, systems, and software — purpose-built for AI workloads like GPU-to-GPU communication at ultra-low latency. Its SkyHammer architecture unifies GPUs, accelerators, memory, and storage into a single fabric, built on SONiC, Ultra Ethernet, and UALink as an open alternative to Nvidia's NVLink. Upscale is fully vertically integrated across the networking stack. It also bets entirely on open standards rather than proprietary protocols, aiming to give hyperscalers and enterprises vendor-neutral interoperability instead of lock-in.
With GPU compute remaining expensive and scarce, organizations are increasingly motivated to maximize the utilization and performance of their existing AI infrastructure, creating strong demand for Upscale AI's optimization solutions.
Upscale has raised over $300M across a $100M+ seed and $200M Series A from Tiger Global, Premji Invest, Intel Capital, Qualcomm Ventures, and others. Its products are targeting hyperscalers and AI infrastructure operators, with networking solutions slated to begin shipping in 2026.
WEKA is a software-defined data platform company that delivers cloud-native, high-performance storage purpose-built for AI and accelerated compute workloads. Its NeuralMesh architecture interconnects data, compute, and AI services across on-prem, cloud, and edge environments, replacing legacy storage with dynamic data pipelines that keep GPUs continuously fed.
WEKA's architecture eliminates the traditional tradeoff between performance and capacity by providing a single namespace that performs like NVMe flash for hot data and scales like cloud storage for large datasets. Their purpose-built data pipeline can feed thousands of GPUs simultaneously without the I/O bottlenecks that plague conventional storage systems.
AI training requires feeding massive datasets to GPU clusters at speeds that traditional storage cannot sustain; WEKA's high-performance data platform solves this by delivering hundreds of gigabytes per second to AI workloads.
WEKA raised $140M in a Series E at a $1.6B valuation, bringing total funding to ~$375M — notably raised entirely from existing investors. The company has over 300 customers including 12 of the Fortune 50 and AI companies like Stability AI, Midjourney, and ElevenLabs, with ARR exceeding $100M and doubling year-over-year.
xLight is building extremely powerful lasers called free electron lasers that produce the specific type of light needed to manufacture the most advanced computer chips.
Today's chipmaking light sources are running up against physical limits, and xLight's technology could provide four times more light while using less energy and fewer consumable materials. Current EUV light sources from ASML use energy-intensive tin plasma lasers that provide only 25% of the light future lithography demands; xLight's FEL approach directly emits EUV light with programmable wavelengths, eliminating consumables like tin and hydrogen. A single xLight source can feed multiple lithography machines, fundamentally changing fab economics.
The surging demand for AI chips is pushing semiconductor manufacturers to their limits, creating urgent need for more productive and cost-efficient lithography solutions that xLight's technology directly addresses.
xLight closed an oversubscribed $40M Series B led by Playground Global in July 2025, and in December 2025 signed a $150M Letter of Intent with the U.S. Department of Commerce under the CHIPS Act — the first award from the Trump Administration's CHIPS R&D Office.
Starcloud is building data centers in space, using satellites equipped with solar panels, radiation shielding, and radiative cooling systems to run GPU compute in orbit. Starcloud launched its first satellite, Starcloud-1, in November 2025, carrying an Nvidia H100 GPU. Their next launch, Starcloud 2, is coming later this year. The satellite will contain GPUs, an AWS server, and a bitcoin mining computer. Starcloud raised a $200M total across a $30M Seed and a $170M Series A in March 2026, the latter led by Benchmark and EQT Ventures, vaulting the company to a $1.1B valuation.
OLIX Computing is building a new class of AI accelerator — the Optical Tensor Processing Unit (OTPU) — that uses photonics to run AI inference workloads faster and more cost-efficiently than GPU-based systems. By integrating SRAM memory directly with optical components, their architecture sidesteps the memory bandwidth bottleneck that limits conventional AI chips without relying on HBM, advanced packaging, or any supply-chain-constrained technology that even the largest hyperscalers struggle to procure.
Unlike GPU designs that depend on expensive high-bandwidth memory stacked alongside the chip, OLIX stores data entirely in on-chip SRAM and uses photonics for interconnect, achieving lower latency and higher throughput per watt at a significantly lower total cost of ownership. Their chips are designed for bit-perfect digital computation — not analog approximation — and are compatible with existing AI models through a custom compiler stack.
Founded in 2024 by 25-year-old James Dacombe, OLIX raised $250M in total funding — including a $220M round led by Hummingbird Ventures at a $1B+ valuation — and has grown to over 70 employees across the UK and North America. The company plans to ship its first OTPU products to customers in 2027.
Efficient Computer designs ultra-low-power general-purpose processors that run AI workloads on devices that can't be plugged into a wall — robots, sensors, satellites, and implantables — without draining their batteries. Their Fabric architecture, a spatial dataflow design developed at Carnegie Mellon University, eliminates the energy overhead of conventional von Neumann processors and delivers up to 100x better efficiency than traditional CPUs, while remaining fully programmable in standard languages.
Unlike fixed-function accelerators that trade flexibility for efficiency, Efficient's approach achieves dramatic power reduction without locking developers into specialized hardware or custom programming models. This makes it practical to run evolving AI software on energy-constrained devices without redesigning the chip every time models change.
Efficient Computer raised $76M in total funding across a $16M seed and a $60M Series A led by Triatomic Capital, with participation from Eclipse, Union Square Ventures, RTX Ventures, and Toyota Ventures.
The Biological Computing Company (TBC) is a San Francisco-based startup, founded in 2022 by neurosurgeon-neuroscientists Alex Ksendzovsky (CEO) and Jon Pomeraniec (COO), the company is building a computing platform that positions biology as a complementary compute substrate — not a replacement for silicon, but a new layer designed to improve the efficiency, stability, and adaptability of existing AI systems.
TBC's core platform encodes real-world data — images, video, and other signals — into living neuronal cultures grown on electrode grids. The neural responses are then decoded into machine-readable representations that are mapped onto frontier AI models through modular adapters. In parallel, the company's Algorithm Discovery platform observes the computational patterns of living neural networks to extract learning principles, using those insights to inform the design of AI architectures beyond the transformer paradigm.
The company reported a 23x retained improvement in video model efficiency using its biological adapters. TBC targets compute-intensive domains including computer vision, generative video, and dynamic world models — application areas where conventional scaling has encountered economic and energy constraints. The underlying thesis is that biological neural networks, refined through billions of years of evolution, offer efficiency and generalization properties that silicon-based systems do not inherently replicate.
TBC emerged from stealth in February 2026 with a $25M seed round led by Primary Venture Partners, with participation from Builders VC, E1 Ventures, Proximity, Refactor Capital, Tusk Ventures, and Wonder Ventures. The company has opened a flagship lab in San Francisco's Mission Bay.
* Primary is an investor in The Biological Computing Company
Nearfield Instruments builds machines that inspect computer chips while they are being made. Its systems drag an atomically sharp probe across the surface of a wafer to map features just billionths of a meter across, like a record needle fine enough to feel the shape of individual transistor structures, reading the wafer without damaging it.
Unlike conventional inspection, which often requires slicing a wafer apart to see inside it, destroying a product worth tens of thousands of dollars to learn whether the process is drifting, Nearfield's QUADRA system measures in full 3D non-destructively and in-line. The company's advance is throughput: an array of miniaturized atomic force microscope heads scanning in parallel, which it claims delivers over 100x the image acquisition speed of conventional single-probe automated AFM tools. Its AUDIRA system extends the same approach to buried features and subsurface voids, which matter increasingly as chips are stacked rather than shrunk.
The structures that AI chips depend on, including high-aspect-ratio DRAM and HBM stacks, gate-all-around and CFET logic, hybrid-bonded 3D packaging, and high-NA EUV resist layers, are precisely the ones conventional metrology struggles to measure, making process control a live constraint on advanced-node yield. QUADRA is qualified and deployed for high-volume manufacturing at leading fabs, and in November 2025 Nearfield signed a multi-year development agreement to install QUADRA at imec's R&D facility in Leuven. In February 2026 the company added a Sidewall Imaging Mode to QUADRA for measuring the flanks of high-aspect-ratio 3D logic and memory structures, unveiled at SPIE Advanced Lithography. The company closed an oversubscribed $380M Series D led by Fidelity in June 2026 at a $1.6B valuation, with Qatar Investment Authority joining alongside Temasek, Walden Catalyst, Innovation Industries, M&G, Invest-NL, TNO Ventures and ING.
VAST Data builds the data foundation that large AI clusters run on. When tens of thousands of GPUs train a model, the bottleneck is often not the chips but getting data to them fast enough; VAST's platform is designed to keep that pipeline saturated, and has since expanded upward into databases, search and agent infrastructure, what the company markets as an AI Operating System.
Unlike conventional storage architectures that force a tradeoff between performance and capacity, VAST's disaggregated shared-everything design separates compute from storage enclosures over NVMe fabric, letting a single namespace behave like fast flash for hot data while scaling to exabytes. Its closest venture comp, WEKA, attacks the same GPU-starvation problem; VAST's differentiation is the move beyond storage into the layers above it, selling database, InsightEngine and AgentEngine components into accounts already anchored on the platform.
The AI datacenter buildout has made the data layer strategically valuable, and VAST has captured an unusual share of the largest deployments. It underpins xAI's Colossus cluster, and CoreWeave signed a $1.17B multi-year agreement in November 2025 making VAST the primary data foundation for its Nvidia-accelerated cloud. The company reports surpassing $4B in cumulative bookings and exiting its last fiscal year above $500M in committed ARR with positive free cash flow. VAST closed a $1B Series F in April 2026 at a $30B valuation, led by Drive Capital with Access Industries co-leading and Nvidia, Fidelity and NEA participating, more than triple its $9.1B Series E mark from late 2023. Since then VAST has added Hyundai Motor Company and Kia as strategic investors in October 2026 after years of production use at their 42dot autonomy unit, expanded its DigitalOcean partnership to run the VAST AI OS inside DigitalOcean's AI-Native Cloud, and launched an agent-sandbox foundation stack with Nvidia at GTC Berlin.
DriveNets sells networking as software. Instead of buying proprietary routers, a carrier or AI datacenter operator runs DriveNets' software across a cluster of standard white-box switches that then behaves as one very large router, the same disaggregation that separated servers from mainframes, applied to the network.
Unlike Nvidia's InfiniBand and NVLink, which deliver high performance inside a single-vendor stack, DriveNets builds its AI fabric on standard Ethernet and positions itself as neutral infrastructure for mixed-vendor clusters, publishing open reference designs with AMD and Broadcom. Its fabric supports scale-up, scale-out and scale-across topologies alongside front-end and storage connectivity, and the pitch is measured in GPU utilization: higher tokens per second and lower cost per token from optimizing the stack end to end.
As operators build clusters spanning multiple accelerator vendors, the networking layer becomes the place where heterogeneity is either enabled or blocked, which is the wedge DriveNets is pressing. The telecom business gives it an unusual profile for a company in this category: cash-flow positive since 2025, with more than $1B in secured business and roughly $250M in trailing revenue. DriveNets closed a $410M Series D on 1 June 2026 led by Bessemer Venture Partners and Atreides Management, with AMD and Red Dot Capital joining as new investors alongside Pitango and D1 Capital, at an $8.5B post-money valuation and $1B raised to date.
Positron builds hardware dedicated to AI inference — the everyday work of answering model queries, rather than the one-time work of training. Its shipping product, Atlas, is a rack-mounted inference appliance that the company says delivers roughly three times the compute per watt of an Nvidia H100 and beats an H200 while drawing about a third of the power. Atlas is fabricated and assembled in the US, and serves requests through an OpenAI-compatible endpoint so customers can swap it in without rewriting their stack.
Unlike accelerators that compete on peak floating-point throughput, Positron designs around memory as the binding constraint. Its next-generation Asimov silicon is specified at up to 2.3TB of memory per chip against 384GB for Nvidia's upcoming Rubin, and the four-chip Titan system is designed to hold 8TB+ of memory, 10M-token context windows, and models up to 16 trillion parameters on a single server. That bet targets long-context LLMs, agentic workflows, and video models — workloads where capacity, not compute, runs out first. Atlas today is FPGA-based; Asimov is the move to custom silicon.
Inference has overtaken training as the dominant share of AI compute spend, and grid power has become the practical ceiling on deployment, making tokens-per-watt the metric customers actually buy on. Positron raised an $875M Series C in September 2026 at a $5B post-money valuation, structured as a $375M Series C co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital and SemiAnalysis Capital, plus a Series C-1 of up to $500M anchored by NEA and Netscape co-founder Jim Clark; NEA's Forest Baskett, Atreides' Gavin Baker, Thomas Jermoluk and Dylan Patel joined the board. The round followed a $230M Series B in February 2026 at a $1B+ valuation co-led by ARENA Private Wealth, Jump Trading and Unless, with Qatar Investment Authority and Arm participating, and brings total funding to roughly $1.18B. More than 50 Atlas racks are deployed, including at Oracle Cloud Infrastructure, serving customers such as Parasail, Jump Trading and i3d.net. Proceeds fund the Asimov tape-out on TSMC N3P by end of 2026, a 2MW+ engineering data center, and the production ramp of Titan, with Asimov production expected in the second half of 2027. CEO Mitesh Agrawal was previously COO of Lambda; co-founder and CTO Thomas Sohmers previously founded REX Computing.