Welcome back to The Compute 100! Today’s edition covers financings and announcements from the last week. We’re eager to be back for a busy fall for the Compute world!
Gimlet Labs raised $300M in a Series B led by Andreessen Horowitz, tripling its valuation to $3B just six months after an $80M Series A. Arm and Microsoft's venture arm, M12, joined as new strategic investors alongside existing backers Menlo Ventures and Factory, bringing Gimlet's total funding to roughly $392M. The company runs a multi-silicon inference cloud: rather than betting on one chip architecture, it routes each inference workload to whichever silicon handles it best — Nvidia GPU, AMD accelerator, or a purpose-built ASIC — and abstracts that fragmentation away from the customer. CEO Zain Asgar said Gimlet is working with Arm specifically to extend that routing layer across Arm's own chip designs. As inference silicon keeps fragmenting across a widening field of general-purpose GPUs and specialized accelerators, the orchestration layer sitting on top of that fragmentation is turning into a chokepoint of its own — one that could end up capturing more durable value than any single chip vendor underneath it, since it earns regardless of which hardware wins the workload. (Bloomberg)
Private Companies
Waferfundraise
Wafer raised $40M in a Series A co-led by Marathon and Chemistry, with AMD Ventures, Wing, and Outset Capital joining — a 50x markup on the $4M seed it raised in April. The company runs autonomous agents that continuously re-tune AI inference deployments in production, and says its optimization layer gets AMD GPUs to roughly 80% of Nvidia B200 performance at under half the cost for some models. Investors are now pricing the software layer that makes non-Nvidia silicon commercially viable, not just the chip vendors themselves. (Wafer)
iPronicsfundraise
iPronics raised $125M in a Series B co-led by Maverick Silicon and Light Street Capital, with Nvidia joining as a new investor, bringing the Spain-based company's total funding to $177M. iPronics builds silicon-photonics optical circuit switches that reconfigure GPU cluster connections in sub-millisecond time, competing with electrical top-of-rack switching as the bottleneck for shuffling traffic between accelerators. Nvidia's direct stake gives the company a foothold inside the GPU maker's own reference architectures for the next generation of AI data centers. (GlobeNewswire)
Crusoefundraise
Crusoe closed a $3B Series F at a $30B valuation, co-led by Atreides Management and Valor Equity Partners with Mubadala Capital participating — triple the $10B valuation it raised at ten months ago — days after signing a five-year, roughly $13B cloud contract with Jane Street, its largest single AI infrastructure commitment yet. The pairing marks Crusoe's shift from Nvidia-backed compute reseller to a scaled AI-cloud provider that growth investors and Wall Street trading desks are now underwriting directly. (TechCrunch)
Nscalefundraise
Nscale is seeking up to $3.5B in financing ahead of a planned US IPO, including as much as $1.5B in convertible notes led by hedge fund Third Point and up to $2B directly from Nvidia, with Goldman Sachs advising and terms benchmarked to a valuation up to $30B. The package follows Nscale's $45B compute deal with Anthropic in August, and the IPO could land within the month. Compute contracts are now routinely large enough to double as IPO collateral before the data centers underlying them are even built. (TechFundingNews)
MatXfundraise
Coatue is in early discussions to form a multibillion-dollar joint venture with MatX to finance the wafers, packaging, and foundry capacity behind the chip startup's push toward 2027 production of its MatX One chip. No deal is signed and talks remain preliminary. The structure would be unusual for a fabless chip startup — financing manufacturing capacity years ahead of volume shipment through a dedicated investment vehicle rather than a conventional equity round. (The Information)
Public Markets
ASML and TSMC announced a joint initiative to shift High-NA EUV lithography from 6-inch to 12-inch photomasks, targeting a 2031 pilot line and removing the die-size stitching limits that constrain today's most advanced scanners. (GlobeNewswire)
Emerging
Compute-in-chiplet: A new paper proposes treating the interconnect between chiplets as compute substrate rather than pure bandwidth infrastructure, pairing RRAM-analog and SRAM-digital compute-in-memory units with a programmable inter-chiplet network. In simulation on Mistral-7B inference, the design claims 2.4x throughput and 27x energy efficiency over Nvidia's H100 — an extraordinary figure from a design-space study with no fabricated silicon and undisclosed benchmark conditions, worth tracking rather than betting on until independent fabrication confirms it. If the architecture holds up in silicon, post-GPU inference stacks could fragment into compute-in-memory chiplet fabrics rather than scaling monolithic dies. (arXiv)
Analog attention: A circuit built in GlobalFoundries' 22nm process computes transformer softmax entirely in the analog domain, skipping the analog-to-digital conversion step that dominates power draw in today's compute-in-memory attention accelerators. The design measures 25.5 picojoules per output element and holds accuracy loss under 2.5% against a digital baseline — a manufacturable path past the ADC bottleneck that has kept analog in-memory accelerators from matching their theoretical efficiency on transformer workloads. (arXiv)