Fractile.ai
Anthropic handed a London chip startup a $250 million purchase commitment for chips that do not exist yet — a bet that signals the frontier AI company’s inference cost problem has grown severe enough to justify early commitment to unproven silicon. The deal, which Bloomberg reported Tuesday, drove Fractile’s pre-money valuation from roughly $1 billion in May to $6.5 billion in advanced funding talks — a jump of more than six times in three months, and the starkest illustration yet of how a single supply agreement can reprice an entire pre-revenue company in the current AI chip market.
Fractile is in advanced talks to raise approximately $600 million, with Redpoint Ventures and Lightspeed Venture Partners set to co-lead, and Thrive Capital and Founders Fund also expected to participate. The round has not closed, and terms could still change. The company’s chips are not expected to be ready for deployment until 2027 — meaning Anthropic is not buying hardware it can put to work today. It is buying priority access to an architectural bet that it believes will help solve one of the central problems in running a frontier AI model at scale: that fetching model weights from memory costs more than computing with them.
Anthropic’s Inference Cost Crisis Made This Deal Rational
To understand why Anthropic would commit $250 million to a 108-person startup that has never commercially shipped a chip, start with the unit economics of running Claude. Anthropic spent an estimated $19 billion on compute in 2026, and its inference costs — the expense of serving Claude’s responses to users in real time — ran 23 percent over budget in 2025, according to reporting by The Information. The company’s gross margins sat at 40 percent in 2025, a significant gap from the 77 percent gross margin target it has told investors it intends to reach before its expected Nasdaq listing.
Inference is where those margin pressures compound. Training a frontier model happens once, over a period of months, on large clusters of GPUs. Inference happens billions of times a day, once per query, every time a user sends Claude a message. The cost of each query is determined primarily by how fast the hardware can move model weights from memory to the compute units that process them — a problem so fundamental that computer architects named it the memory wall in 1994. For a large language model like the ones Anthropic deploys, generating a single response requires repeatedly streaming hundreds of gigabytes of model parameters from off-chip memory — a process that, at Anthropic’s scale, compounds into a structurally expensive overhead that no amount of additional GPU purchasing resolves.
Custom inference chips from established hyperscalers — Google’s Tensor Processing Units, Amazon’s Trainium, Microsoft’s Maia — all address this problem within the GPU architectural paradigm: faster interconnects, denser High Bandwidth Memory stacks, more efficient scheduling. What they do not do is eliminate the data movement itself. Fractile’s memory-compute fusion approach claims to do exactly that.
What Memory-Compute Fusion Actually Means
Fractile calls its approach memory-compute fusion — and the name captures the architectural distinction precisely. Standard inference chips, including Nvidia GPUs with High Bandwidth Memory, separate the memory where model weights live from the compute units where matrix multiplications happen. Moving data between those two locations is the bottleneck: for a 70-billion-parameter model at 16-bit precision, a GPU must transfer roughly 140 gigabytes of model weights to generate a single token. At any production scale, that movement dominates both latency and energy consumption.
Fractile’s design eliminates the movement. The company has licensed the Andes AX45MPV RISC-V vector processor — a 64-bit, in-order, dual-issue core with a 1,024-bit Vector Processing Unit and high-bandwidth vector local memory — and is incorporating it into a data center AI inference accelerator that executes 99.99 percent of the operations needed to run model inference directly in on-chip SRAM. The custom extension tooling — Andes ACE (Automated Custom Extension) — allows Fractile to add specialized vector and scalar instructions tuned specifically to the arithmetic patterns of transformer inference, rather than relying on a generic instruction set.
The phrase “99.99 percent of operations in on-chip memory” is the key claim. By baking the computational operations into the memory array rather than shuttling parameters back and forth, Fractile’s architecture targets the decode phase of transformer inference directly. Decode — the process of generating each token sequentially after the initial prompt has been processed — is not compute-bound. It is memory-bound. The chip’s compute units sit idle most of the time, waiting for data. An architecture that eliminates that wait addresses the actual bottleneck rather than adding more waiting capacity.
The company claims this approach can run large language models up to 100 times faster than existing hardware while reducing operational costs by as much as 90 percent. These are simulation-derived figures from a company that has not yet shipped a commercial chip. No independent benchmark organization has verified them under production conditions. Walter Goodwin, Fractile’s CEO and co-founder — an Oxford PhD who completed his doctorate in robotics at the university’s Robotics Institute — said in a February 2025 interview that the company had multiple test chips planned and a tape-out imminent at that time, with teams working across London and Bristol. Whether those test chips have validated the simulation-based performance claims is not publicly known.
Why Anthropic Made a $250M Bet Before the Chips Exist
Anthropic now has, or is pursuing, relationships with an unusual number of chip suppliers simultaneously: Nvidia GPUs (existing), Google Tensor Processing Units under a deal that includes 3.5 gigawatts of TPU compute from 2027, Amazon Trainium through Project Rainier, the reported Microsoft Maia 200 and Maia 300 discussions, an in-house chip design program announced August 5, exploratory Samsung foundry talks, and now Fractile.
The multi-supplier posture is not redundancy for its own sake. Each relationship covers a different architectural approach to the inference cost problem, and Anthropic’s publicly stated strategy — matching workloads to the chips best suited for them — requires options across the architectural spectrum. What Fractile offers that none of the others does is a genuine departure from the HBM-dependent memory hierarchy: no High Bandwidth Memory stacks, no off-chip DRAM movement, compute integrated directly into SRAM. If the architecture proves out at production scale in 2027, it would give Anthropic a chip tier with qualitatively lower inference cost than any current HBM-dependent alternative.
The chip industry operates on long lead times, and customers that want priority access to novel silicon must lock in agreements well before production begins. That is the structural logic of the Anthropic deal: a $250 million initial commitment buys Fractile the capital it needs to complete its tape-out and production ramp, while giving Anthropic a guaranteed position in the delivery queue for a chip it believes could reduce inference costs more substantially than any incremental improvement to existing hardware. The 2027 delivery timeline means Anthropic is not buying hardware today — it is buying an option on a different cost structure.
Three Architectural Bets, One UK Ecosystem
Fractile is the third distinct inference-chip architecture to attract major UK-based and US venture capital in the current cycle, and the comparison illuminates the different engineering approaches the market is backing simultaneously.
Etched, whose Sohu chip shipped its first rack to quantitative trading firm Jane Street last week at a $21 billion valuation, hardwires transformer computation directly into silicon — a fixed-function approach that eliminates the general-purpose overhead of GPU instruction scheduling (which leaves 60 to 70 percent of a GPU’s theoretical compute capacity idle on transformer workloads) but cannot be reprogrammed for non-transformer architectures. Olix, the London startup that raised $312 million at a $3.3 billion valuation earlier this month, uses silicon photonic interconnects to bypass HBM entirely through a different mechanism — high-bandwidth optical data movement rather than computation-in-memory.
Fractile’s approach — true processing-in-memory using RISC-V vector cores embedded in SRAM — is the most architecturally radical of the three. Etched eliminates CUDA overhead; Olix eliminates HBM bandwidth constraints; Fractile claims to eliminate memory data movement itself. Groq, a longer-established inference startup that Nvidia licensed in a roughly $20 billion deal, uses a comparable SRAM-centric philosophy but does not claim 99.99 percent on-chip operations — it keeps data movement within a deterministic software-defined schedule rather than eliminating it entirely.
The UK has now produced three distinct inference chip startups with billion-dollar-plus valuations in the same 12-month window: Fractile (implicitly valued at $6.5 billion in the new round, if it closes), Olix ($3.3 billion), and a broader ecosystem that includes the UK Sovereign AI Fund, which launched in April 2026 with a £500 million (approximately $677 million) commitment. That concentration in a single country’s early-stage chip ecosystem does not occur by accident — it reflects a combination of deep-tech academic talent (Oxford, Cambridge, Bristol, Imperial College) and early-stage venture capital willing to back semiconductor bets that US investors historically underweighted outside of Silicon Valley.
What the Six-Times Valuation Jump Reveals
The $1 billion to $6.5 billion repricing in 90 days is extreme even by the inflated standards of 2026 AI venture capital. It is also, in venture arithmetic, not irrational.
A $250 million customer contract from one of the three largest frontier AI labs, committed before production — with stated intention to expand — converts a startup’s entire premise from “interesting architectural thesis” into “named customer validation.” The AI chip market is not a commodity market where many suppliers serve interchangeable buyers. It is a market where a single anchor customer defines the supply chain for years. Anthropic’s commitment does not guarantee that Fractile’s chips will perform as claimed, or that the production ramp will succeed, or that the architecture will remain competitive with what Nvidia and other incumbents ship in 2027. What it does is validate the company’s technical credibility with the counterparty most capable of performing the relevant engineering due diligence.
Some of the money in the new round was invested at a lower valuation — Bloomberg noted this explicitly — meaning the headline $6.5 billion figure reflects top-end pricing, not a uniform clearing price across the entire raise. The round has not yet closed and terms could still change.
Execution Is the Only Question That Matters Now
Fractile’s elevation to the front rank of AI chip startups comes with a corresponding set of execution risks that the valuation does not price away.
Tape-out — the process of finalizing a chip design for manufacturing — is a genuine technical gate. Cerebras Systems, now publicly traded, disclosed tape-out delays in filings — stating it has experienced and may continue to experience delays in securing tape-out slots and resolving technical issues with new designs, and that failures at tape-out can require restarts of the design cycle. The chip industry has a documented history of well-funded startups that designed compelling silicon but could not execute the manufacturing ramp.
Fractile’s central performance claims — 100x speed, 90% cost reduction — are derived from simulations, not production measurements. The claim that 99.99 percent of operations execute on-chip presupposes that model weights can be held in the available SRAM capacity. Large frontier language models can weigh hundreds of gigabytes; on-chip SRAM is expensive per bit and limited in area. How Fractile resolves this constraint — through aggressive quantization, mixture-of-experts sparsity exploitation (which Goodwin noted in a February 2025 interview as a key design signal from DeepSeek’s architecture), or hierarchical tiling — is not publicly documented in production detail.
The CUDA ecosystem switching cost is real. Moving to any non-Nvidia inference architecture requires rebuilding the production inference software stack from scratch. Anthropic’s multi-chip strategy, which explicitly includes Nvidia GPUs, suggests the company is not betting on a single architectural winner — it is hedging across approaches and will route workloads to the hardware that proves most efficient for each class of query.
Fractile also navigated an early governance challenge. Co-founder and original CTO Yuhang Song departed in May 2024 after questions arose about his prior academic ties to Beihang University — one of China’s Seven Sons of National Defence, institutions with close research relationships with the People’s Liberation Army, according to reporting by Sifted and City AM. There is no suggestion of wrongdoing by Song; the departure addressed the security concern proactively, consistent with Fractile’s early investment from the NATO Innovation Fund. Song is now an associate professor at Nanjing University’s School of Artificial Intelligence.
In-Memory Computing’s Biggest Validation Yet
The largest implication of this deal is one the funding numbers obscure: Anthropic’s $250 million commitment is, in effect, the most significant institutional endorsement that processing-in-memory computing has ever received as a production architectural approach.
In-memory computing — the idea of co-locating compute and memory to eliminate the von Neumann bottleneck — has been a research priority in computer architecture for three decades. Commercial deployments exist (Groq, Cerebras) but have not historically attracted frontier AI lab supply commitments before production. Anthropic’s willingness to commit $250 million before Fractile has shipped a chip is a statement that its engineering teams, after evaluating the simulation data and the architecture in technical depth, believe the approach is viable at production scale. That assessment — from a company spending approximately $19 billion per year on compute infrastructure and deeply motivated to find inference cost reductions — carries more evidential weight than any benchmark Fractile could publish.
Whether the chips ship on time, perform to specification, and prove competitive against what Nvidia, Google, and Amazon will be offering in 2027 are the questions that will ultimately determine whether $6.5 billion was a prescient bet or a peak-market artifact. But the architectural bet itself — that eliminating memory data movement is a viable path to dramatically cheaper inference — has now received its most credible endorsement.
Currency conversions are approximate, based on rates at time of publication.
Frequently Asked QuestionsWhat exactly did Anthropic agree to with Fractile, and why does it matter before the chips exist?
Anthropic signed an initial agreement to purchase approximately $250 million worth of Fractile’s inference chips, with stated intention to expand the contract. The chips are not expected to be ready until 2027. The deal matters before delivery because it converts Fractile’s architectural thesis into a named-customer commitment from one of the three largest frontier AI labs — the counterparty most capable of performing serious engineering due diligence on an unproven architecture. At Anthropic’s compute scale ($19 billion in estimated annual spending), even a 30 to 40 percent inference cost reduction from a new architecture translates into billions of dollars annually. The $250 million commitment is the price of a position in the queue for that potential savings.
How does memory-compute fusion differ from what Nvidia GPUs and other inference chips do?
A standard GPU — including Nvidia’s H100 with High Bandwidth Memory — keeps model weights in a separate off-chip memory array and moves them to compute cores when needed. For a 70-billion-parameter language model, this means transferring roughly 140 gigabytes of data to generate each token, repeatedly, because memory bandwidth is shared between the model weights and the intermediate data (the key-value cache) needed during generation. Fractile’s architecture claims to eliminate this movement by integrating computation directly into on-chip SRAM, executing 99.99 percent of inference operations without accessing off-chip memory. Etched’s Sohu chip addresses the same problem differently — by hardwiring transformer operations into silicon — while Olix uses silicon photonic interconnects to move data faster rather than eliminate its movement. Fractile’s approach is architecturally the most radical of the three: it targets the data movement cost directly rather than accelerating or reducing it.
What are the real risks that Fractile’s $6.5B valuation doesn’t account for?
The performance claims (100x speed, 90% cost reduction) are derived from simulations of a chip that has not yet been manufactured at production scale. Chips differ from simulations in yield rates, thermal behavior under load, and performance across the full distribution of production workloads — not just the optimized test cases used in benchmarks. Frontier language models weigh hundreds of gigabytes; whether Fractile’s on-chip SRAM architecture can accommodate those model sizes without a significant engineering workaround is not publicly documented. Tape-out failures and manufacturing delays are real risks for any chip startup; Cerebras disclosed these as material risks in its SEC filings. The 2027 delivery timeline also means that Nvidia, Google, and Amazon will have new hardware generations in the market by the time Fractile delivers — and what those alternatives offer will determine whether Fractile’s cost advantage holds.
Why is the inference cost problem so important to Anthropic right now?
Anthropic runs Claude for millions of enterprise and consumer users daily. Inference — serving each query — is not a one-time cost like training; it is a continuous, per-query expense that scales with usage. The company’s annualized revenue run rate reached $30 billion in early 2026, according to SemiAnalysis data, but inference costs ran 23 percent over budget in 2025. Gross margins at 40 percent are well below the 77 percent target Anthropic has stated for its expected IPO, and the gap is driven primarily by compute costs on the inference side. Every architectural approach that reduces per-token inference cost — whether through custom silicon, software optimization, or in-memory computing — directly expands gross margins without requiring additional revenue. That is why Anthropic is simultaneously pursuing Trainium, TPUs, Maia negotiations, in-house chip design, software optimization through an attempted Decart acquisition, and now Fractile: each is a bet on a different path to the same economic destination.