On November 11, 2026, a four-year-old British chip startup raised $720 million in Series C funding at a $7 billion valuation. Fractile is not building another training GPU. It is betting that the next bottleneck in AI is how fast models produce tokens, and that the fix is stacking DRAM directly on top of logic.

That bet puts Fractile in the same conversation as Groq, Etched, Cerebras and a long list of inference specialists, but with a different memory story. Nvidia still dominates the racks. Fractile says the memory wall will decide who wins the next wave.

This guide covers what Fractile claims to build, how the stacked-DRAM architecture differs from HBM and SRAM designs, the funding trail from seed to Series C, how it stacks up against other inference startups, and where the public evidence is still thin. Last updated: November 19, 2026.

Key Takeaways

Fractile at a glance

Here's the short version. Treat performance claims as vendor and investor language until silicon ships and third parties measure it.

ItemDetail
Founded2022 (summer), London and Bristol
FoundersWalter Goodwin (CEO); Yuhang Song (co-founder, per EU-Startups)
HeadquartersLondon, with Bristol hardware operations
FocusInference accelerators and rack-scale systems
Architecture claimStacked DRAM on logic die
Series C$720 million; $7 billion post-money (Nov 11, 2026)
Series B$220 million (May 2026)
SeedAbout $15 million / €13 million (2024)
First commercial chipsTargeted for 2027
UK public capitalSovereign AI Venture Fund and British Business Bank

Source: Data Center Dynamics on Series C size and valuation; GOV.UK on Sovereign AI and British Business Bank participation; EU-Startups on founders, seed and Series B euro figures; TechCrunch on 2027 ship timing; Lightspeed on stacked DRAM.

💡
What Fractile sells, in one line: purpose-built AI inference chips and systems that put DRAM on top of the compute die, so large models can generate tokens faster and with longer context than today's HBM-bound GPUs. It does not sell training GPUs, and as of November 2026 it has not shipped a commercial chip.

Why Fractile exists: the inference time bottleneck

Fractile's founding thesis is blunt. Capability is no longer the only limit. Time-to-output is. The company's manifesto says today's LLMs already produce up to 100 million tokens on hard problems, and at roughly 40 tokens per second that can take about a month.

To shrink a month of generation into a day, Fractile says you need about 1,200 tokens per second while still holding huge contexts. That is the product problem: speed and capacity together, not one or the other.

Lightspeed, a Series C co-lead, frames the market the same way. Its investment note says frontier models are processing more than 10x more tokens every year, and that over the next five years roughly 80% of AI chip spend will go to inference.

Here's why that matters for buyers. Training clusters get the headlines. Inference is where token costs, latency SLAs and agentic workloads collide. Inside Deep Tech's AI inference chips comparison maps the broader field Fractile is entering.

The memory wall Fractile is trying to break

During autoregressive generation, a chip spends most of its time waiting on weights and KV cache. Lightspeed puts it simply: every token requires reading weights and context out of memory, so the speed a single user feels is set by memory bandwidth.

HBM raised bandwidth by stacking DRAM beside the GPU over a silicon interposer. SRAM-heavy designs keep hot data next to compute but run out of capacity for long contexts. Fractile's pitch is to take the DRAM that normally goes into HBM stacks and package it directly on top of the logic, chasing both bandwidth and capacity on one chip.

CEO Walter Goodwin told LinkedIn, as quoted by TechCrunch and DCD, that Fractile has already extracted "20x higher speeds from DRAM memory." That is a company claim, not an independent benchmark.

If you need the HBM side of this trade-off, Inside Deep Tech's HBM full guide and SOCAMM LPDDR5X guide cover the memory products already shipping into AI servers.

The agentic angle makes the trade-off sharper. TechCrunch notes that longer task chains with large models amplify the memory-capacity versus bandwidth tension. SRAM designs stay fast until context blows up. HBM designs hold more weights but still serialize a lot of traffic off-die. Fractile is selling a third point on that curve.

What Fractile says it is building

Public product pages are thin. Fractile's homepage describes chips and systems that generate thousands of tokens per second on large models with very long contexts. It does not publish a named SKU, TDP, process node or rack drawing.

What is public:

  • Inference-only accelerators. Fractile is a chip-first company aimed at serving models, not training them, per its homepage and Lightspeed note.
  • Rack-scale systems. Lightspeed says Fractile is designing accelerators and the racks around them, not only a chiplet sold into someone else's board.
  • Full-stack ownership. Lightspeed says Fractile owns the stack instead of handing backend design to a Broadcom-style partner, which is meant to control cost and architecture.
  • Compiler and ML research. TechCrunch notes a machine-learning compiler alongside the silicon, so software is part of the performance story.
  • Hiring footprint. The manifesto lists London, Bristol, Palo Alto / Bay Area and Taipei as hiring locations.

What is not public: a datasheet, a published TOPS or tokens-per-second number measured by a customer, a foundry partner name, or a confirmed process node. Until those appear, compare Fractile on architecture thesis and capital, not on a scoreboard.

That thin public surface is normal for a fabless AI chip company before first silicon. It is also a diligence flag. Buyers should ask for a roadmap that separates silicon bring-up, software maturity and rack integration, because Lightspeed's rack-scale framing means Fractile is not only a chiplet vendor.

Funding trail: seed to a $7 billion Series C

Fractile's capital story moved fast in 2026. The numbers below are the ones that appear in primary or near-primary reporting.

RoundDateAmountValuationLeads / notable backers
Seed2024About $15M (€13M)Not disclosedEarly venture backers (per EU-Startups)
Series BMay 2026$220MAbout $1B post-money (reported)Accel, Factorial Funds, Founders Fund
Series CNov 11, 2026$720M$7B post-money / over £5BThrive, Founders Fund, Lightspeed, Redpoint; Sovereign AI; BBB

Source: EU-Startups on seed and Series B euro figures; Latham & Watkins and Accel on Series B leads; Silicon UK on the reported $1 billion Series B valuation; DCD and TechCrunch on Series C size and $7 billion; GOV.UK on the over-£5-billion UK figure and public investors.

Series C participants named across DCD, TechCrunch and Walter Goodwin's LinkedIn post (via those outlets) include Dragoneer, DST Global, BlackRock, T. Rowe Price, CapitalG, Accel, Conviction, Felicis, Balderton, Shopify Ventures and others. Exact checks per firm are not public.

Nine months before the Series C, DCD notes Fractile planned to invest £100 million ($135 million) over three years into UK sites, including a new Bristol hardware engineering facility. That expansion plan is separate from the Series C cash, but it shows where the company wants headcount.

Series B context still matters. Accel co-led the May 2026 round and described Fractile as rethinking compute, memory bandwidth and software together. Latham & Watkins lists Conviction, Gigascale, O1A, Felicis, Buckley Ventures and 8VC among participants. That syndicate already looked like a late-stage AI hardware roster before the November jump to $7 billion.

⚠️
Read the valuation carefully: the UK government says over £5 billion. Trade press consistently reports $7 billion. Those are the same round described in different currencies, not two separate raises.

How Fractile compares with other inference bets

Fractile sits in a crowded inference field. The table below is a map of approaches, not a performance ranking. Only shipping products have independent scores.

CompanyApproachShip status (Nov 2026)Capital markerHonest difference
FractileStacked DRAM on logic; full-stack inferenceFirst chips targeted 2027$720M Series C; $7B valuationMemory packaging thesis; no public silicon yet
GroqLPU / SRAM-centric inferenceCloud and on-prem systems shippingAcquired path into Nvidia orbit for LPX lineReal deployments; different memory trade-off
EtchedASIC for transformer decodePre-volume / early silicon narrativeHigh-profile private raiseModel-specific silicon vs general DRAM stack
CerebrasWafer-scale engineCS-3 systems shippingPublic traction via cloud and enterprisesWafer-scale vs stacked DRAM

Source: Fractile Series C coverage and 2027 timing; Inside Deep Tech's Groq LPU guide, Gavin Uberti / Etched founder guide and Cerebras WSE-CS3 guide for the peer baselines.

The useful comparison is not "who is fastest on a slide." It is who can prove tokens per dollar, tokens per watt and long-context capacity on production models. Groq and Cerebras already face that test in the field. Fractile does not yet.

For cost-focused buyers, Inside Deep Tech's real cost of running AI in 2026 breaks down how inference chip choice shows up in unit economics.

Why the UK government cares

Fractile is also a sovereign-capability story. The GOV.UK release frames the Series C as part of reducing dependence on overseas AI hardware supply chains and keeping chip design jobs in Britain.

AI Minister Kanishka Narayan and Business Secretary Jonathan Reynolds both appear on the record in that release. Sovereign AI Investment Partner Joséphine Kant and British Business Bank CIO Leandros Kalisperas praise the full-stack chip-and-software approach.

The same release notes the Budget added further backing for Sovereign AI, including £140 million more for the fund family (£60 million for investments plus £80 million for an R&D procurement scheme), and expands Sovereign AI into Physical and Industrial AI. Those are government program figures, not Fractile-specific checks.

For investors and operators outside the UK, the point is simpler: Fractile now has both top-tier US venture capital and UK public capital on the cap table. That helps hiring and narrative. It does not ship wafers.

EU-Startups places Fractile in a wider 2026 European AI silicon wave that also includes OLIX, EUCLYD and others. Fractile is the largest disclosed round in that set, which raises the execution bar rather than lowering it.

What Fractile means if you buy or build AI infrastructure

You will not rack a Fractile system tomorrow. You can still decide whether to track it as a 2027 to 2028 option.

  • If you run agentic or long-reasoning workloads: Fractile's thesis matches your pain. Long token chains punish memory bandwidth. Ask for measured tokens/sec on your model family, not a generic claim.
  • If you are GPU-locked for training: Fractile is an inference add-on story, not a training replacement. Keep Nvidia or AMD training spend separate from any Fractile pilot budget.
  • If you evaluate startups for rack design: insist on power, cooling, networking and software stack details before a PoC. The Lightspeed note says Fractile builds rack-scale systems, so the integration surface is large.
  • If you care about UK compute sovereignty: Fractile is one of the few domestic AI silicon names with a multi-billion valuation. Track Bristol and London hiring as a leading indicator of execution.

Honest limits: where the Fractile story is thin

The raise is real. The product proof is mostly still ahead. Weigh these gaps before treating Fractile as a near-term vendor.

  • No commercial silicon yet. Goodwin says first chips and systems ship next year. Until then, every performance claim is forward-looking.
  • No public MLPerf or customer benchmark. Unlike shipping inference vendors, Fractile has not posted a third-party score as of November 19, 2026.
  • No named hyperscaler customer. Coverage names investors and government backers, not production buyers.
  • Architecture details are sparse. Stacked DRAM on logic is clear at the slogan level. Foundry, node, HBM supplier displacement and yield story are not.
  • Valuation assumes a hard landing in 2027. A $7 billion price for a pre-revenue chip company leaves little room for slip. Etched, Groq and Cerebras already compete for the same buyer attention.
  • Earlier pivot language. TechCrunch says Fractile's DRAM focus came "after a pivot," which means the current architecture is not the company's first bet.

Inside Deep Tech's take

🧭
Inside Deep Tech's take: Fractile just became one of Europe's most important AI chip stories overnight, and the stacked-DRAM thesis is coherent for agentic inference. The honest bar is still 2027 silicon that beats HBM GPUs on tokens per watt and long-context serving cost for a named model. Capital does not clear that bar. If Fractile ships on time with independent numbers, the $7 billion valuation will look early rather than crazy. If the first systems slip into 2028 without customer logos, Fractile risks becoming another well-funded inference architecture that never leaves the demo rack.

FAQ

What is Fractile?

Fractile is a UK AI hardware company founded in 2022 that builds inference accelerators and systems using stacked DRAM on logic. It is private, based in London and Bristol.

How much funding has Fractile raised?

The disclosed path includes about $15 million seed, a $220 million Series B in May 2026, and a $720 million Series C in November 2026. Exact cumulative total depends on undisclosed earlier checks.

What is Fractile's valuation?

Trade press reports a $7 billion post-money valuation on the Series C. The UK government describes the company as valued at over £5 billion.

When will Fractile chips ship?

CEO Walter Goodwin says the Series C will fund first chips and systems next year, meaning 2027. No public production ship date with customer names is confirmed.

How is Fractile different from Groq or Etched?

Groq ships SRAM-centric LPUs today, covered in Inside Deep Tech's Groq guide. Etched targets transformer-specific ASICs, covered in the Gavin Uberti guide. Fractile's public differentiator is stacking DRAM on the logic die for bandwidth and capacity together.

Does Fractile compete with Nvidia?

On inference serving, yes in ambition. Nvidia still owns most training and a large share of inference GPUs. Fractile is pitching a specialized alternative, not a drop-in CUDA clone, and has not published a customer displacement case.

What to watch next

  • A named process node, foundry or packaging partner tied to the 2027 chip.
  • First customer logos or cloud availability windows after silicon arrives.
  • Independent tokens/sec and tokens/watt numbers on a frontier open model.
  • Whether Bristol's hardware facility, flagged by DCD, shows up in hiring and supply-chain announcements.
  • How Fractile shows up in the next refresh of Inside Deep Tech's AI inference chips comparison.

Fractile's next chapter is about silicon, not press releases. If you are planning 2027 inference capacity, keep Fractile on the watchlist and demand measured data before you reserve rack space. Inside Deep Tech's best AI inference chips of 2026 is the best place to compare what already ships.

High-Bandwidth Memory (HBM): A Full Guide
How HBM stacks, supply and packaging constrain AI accelerators today.

Cover image: generated semiconductor wafer and stacked memory dies under blue lab light, blurred with title overlay for this post.