영국 AI 추론 칩 스타트업 Fractile가 시리즈 C에서 7억 2천만 달러를 조달했으며, 보고된 기업가치는 70억 달러다. 아직 출시된 칩은 없다. 영국 정부의 Sovereign AI 펀드와 British Business Bank가 이번 투자에 참여했다.

Fractile, the London-based AI inference chip startup, announced a $720 million Series C on November 11. Its reported valuation rests on one unproven engineering bet: placing DRAM directly on top of logic, stacked so that AI models can generate answers faster. The company has not shipped a production chip. The UK government's Sovereign AI Fund is among the backers.

For anyone tracking AI hardware, the round offers two things: a concrete picture of how Fractile's memory design differs from GPUs and SRAM-based rivals, and a clear list of what the company has not yet shown. Reuters, citing a person involved, reported a $7 billion valuation on Wednesday. CEO Walter Goodwin said the funding will let Fractile ship its first chips and systems in 2027.

**Four Venture Firms Led the Round While the UK Sovereign AI Fund Took Part**

Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint led the round. New investors include CapitalG, BlackRock, DST Global, Dragoneer, Elad Gil, Balderton, SV Angel, the British Business Bank and the Sovereign AI Fund. Returning investors include Accel, Conviction, Factorial and Felicis. Founders Fund also led the $220 million Series B in May, alongside Accel and Factorial Funds.

Fractile lists offices in London, Bristol, Palo Alto and Taipei and is hiring across its stack.

**Government, Financial Times and Reuters Report Different Valuations for the Same Round**

The UK government says the round values Fractile at over £5 billion (approximately $6.6 billion), according to TechCrunch's report on the round. The Financial Times and a person cited by Reuters put the figure at around $7 billion, according to the same report, which hedged that number. Some outlets still list $6.5 billion, which Bloomberg reported in August as a pre-money figure, meaning the value before new capital is added. Adding $720 million of new capital to $6.5 billion gives about $7.2 billion, close to Reuters' figure.

The step-up is steep. Fractile raised a $220 million Series B in May, valuing it near $1 billion, after a seed round of about $15 million in 2024. The new valuation is more than six times the May figure.

The size of the round also changed as it came together. Bloomberg reported in August that Fractile was in talks to raise about $600 million at a $6.5 billion pre-money valuation, with some of the money invested at lower prices. Companies House filings show about £577 million (approximately $765 million) of Series C shares allotted between September 9 and October 23, Tech Funding News reported, in three share classes priced at £801.54, £707.25 and £1,452.35 per share. Sifted reported £540 million (approximately $716 million).

The coverage reviewed does not explain the gap, but the arithmetic is suggestive. Dealroom reports that the £37.4 million C-2 class came from earlier investments converting into shares. Subtracting it from £577 million leaves about £540 million. The announced $720 million equals roughly £543 million at today's rate. Because shares sold at prices that differ by about a factor of two, a single headline valuation depends on which class sets it.

**How Does Stacking DRAM on Logic Differ From HBM and SRAM?**

AI inference, the stage where a trained model produces answers, is limited mostly by memory bandwidth rather than arithmetic. Each generated token requires streaming the model's weights from memory to the processor, so the chip that moves weights fastest produces tokens fastest. Dealroom's summary of an October No Priors interview says Goodwin argued that each new model demands far more memory bandwidth and that data-center inference economics come down to cost per gigabyte.

Chip designers have three memory options, and each involves a trade-off. SRAM sits on the die and is fast but small. Nvidia's Groq 3 LPU holds 500 MB at 150 TB/s, while a Rubin GPU holds 288 GB of HBM4 at 22 TB/s, so a 256-chip LPU rack holds only 128 GB in total. An SRAM cell needs four to six transistors per bit, while DRAM needs one transistor and one capacitor, which is why DRAM is denser and cheaper per bit. HBM stacks up to 32 DRAM dies linked by through-silicon vias beside the processor. HBM4 reaches up to 2 TB/s and 64 GB per stack.

Fractile's earlier design, as VKTR described it in August, put compute and memory on one die using SRAM. Goodwin said the company concluded by late 2023 that SRAM could not scale, and it pivoted to extracting high bandwidth from low-cost DRAM in work with memory vendors. The company now places DRAM directly on top of logic, aiming for more bandwidth than HBM and more capacity than SRAM. Coverage as recent as August still describes the chip as SRAM-based, a description that conflicts with the DRAM design Fractile now describes.

Academic work on DRAM-on-logic stacks shows what is at stake. A 2025 arXiv paper on Stratum puts HBM's through-silicon via pitch at about 10 µm (0.0004 inches), compared with about 1 µm (0.00004 inches) for copper-to-copper hybrid bonding, which allows far more connections between memory and logic. A 2026 evaluation framework paper cites access energy of 0.66 to 0.88 picojoules per bit for 3D-stacked DRAM, 77% to 83% below HBM, a figure it takes from another source. These are estimates for the technology class, not Fractile measurements. Fractile has not said which bonding process it uses.

**Thermal Limits, Yield and Packaging Remain Open Engineering Questions**

Stacking brings its own limits. In 3D chips, heat builds up inside the stack and hotspots need careful management. In wafer-to-wafer bonding, a single defective die can ruin an entire stack. Samsung's zHBM concept, which stacks DRAM directly on a processor, targets roughly four layers rather than the 12 or 16 of conventional HBM, because of heat. A 2026 DeepStack modeling study found that, at 7 nm, effective bandwidth peaks below about 10 stacked DRAM layers and then declines. The coverage reviewed for this article includes no Fractile figures on layer count, bandwidth, power, thermal behavior or yield.

Manufacturing runs through Taiwan. A Fractile job listing for a Taiwan-based packaging engineer describes owning packaging yield and ramp while working with outsourced assembly-and-test providers and substrate suppliers.

**Tight DRAM Supply Could Test the Premise That Commodity Memory Is Cheap Capacity**

Fractile's pivot depends on DRAM being a cheap, abundant substrate, and the 2026 memory market does not look like that. Wikipedia's DRAM entry reports early-2026 price increases exceeding 200% for some DDR and flash memory since early 2025. Its HBM entry reports HBM crowding out commodity DRAM, with Micron citing a 3-to-1 wafer conversion ratio between HBM and DDR5. Tom's Hardware reported in August that memory makers shifted capacity toward HBM and server DRAM, starving commodity markets. Goodwin has said Fractile works with memory vendors, but whether that insulates it from market swings is not established. The company has not said who supplies its DRAM dies or on what terms, so the exposure is an open question rather than a documented problem.

**Performance Claims Range From 25 Times to 100 Times Faster, With No Public Benchmarks**

Fractile has described its speed advantage differently over time. At its 2024 debut, it said its design could run large models up to 100 times faster and 10 times cheaper than existing hardware. Computer Weekly reported the company's claim as up to 50 times faster at about 10% of the cost. In more recent investor materials, the framing is reportedly 25 times faster at one-tenth the cost. The company's May announcement targeted throughput rising from 40 tokens per second to more than 1,000, W.Media reported. The coverage reviewed includes no public benchmark data.

Goodwin has argued that fast inference matters most for AI agents, with a goal of multi-trillion-parameter models running at many thousands of tokens per second, according to Dealroom's summary of the October interview.

**UK Sovereign AI Fund Joined the Round but Has Not Disclosed Its Stake**

The UK Sovereign AI Fund was announced in April with £500 million (approximately $663 million) of public money and invests directly in British AI companies. The Telegraph reported that its checks typically run from £1 million (approximately $1.3 million) to £10 million (approximately $13.3 million), and that it entered at a slightly lower valuation than the final price, according to Dealroom's summary. The size of its stake has not been disclosed, and Fractile and the government declined to comment. The British Business Bank also took part in the round.

The government said it wants Britain to hold a larger stake in the hardware underpinning its future economy, TechCrunch's report on the round noted. Fractile had already announced a £100 million UK expansion (approximately $133 million) in February, and Computer Weekly reported that the plan covers its London and Bristol sites.

The public money goes into