SambaNova's $1bn Series F: An $11bn Bet Against Nvidia

What changed

• SambaNova completed the first close of a Series F round on 8 July 2026 — $1bn in strategic financing at an $11bn post-money valuation, led by General Atlantic.
• It is explicitly a "first close." SambaNova and General Atlantic both say additional investors are expected to join in the coming weeks as a second close completes, meaning the final round size could grow beyond $1bn.
• The backer list is unusually institutional for an AI-chip round: General Atlantic led, with Seligman Ventures, T. Rowe Price Associates, Capital Group, BlackRock, Battery Ventures, Intel Capital, Qatar Investment Authority (QIA) and Vista Equity Partners among the names disclosed.
• The valuation has swung hard in seven months. Reports place an Intel acquisition offer at roughly $1.6bn as recently as December 2025; SambaNova's $11bn tag today is about 7x that figure.
• The pitch behind the cheque: SambaNova's Reconfigurable Dataflow Unit (RDU) chips process inference differently from a GPU, and General Atlantic is betting that difference matters more as inference — not training — becomes AI's larger, stickier compute market.

The round: from a stalled Intel deal to $11bn in seven months

Who's actually in the cap table

SambaNova is a Palo Alto-based AI systems company founded in 2017 by veterans of Sun/Oracle and Stanford, and it has spent the last two years repositioning itself away from training hardware and toward enterprise inference. The Series F first close was announced jointly by General Atlantic and covered independently by TechCrunch and Bloomberg, with the same round size, valuation and lead investor confirmed across all three. SambaNova CEO Rodrigo Liang framed the raise around one line: "inference has become foundational to enterprise and industry transformation." General Atlantic's Martin Escobari put the demand case more bluntly, saying demand for inference capacity is "accelerating well ahead of supply." The company says proceeds go toward expanding capacity, accelerating product development and scaling deployments globally — not a pivot into a new business line.

The valuation whiplash, explained

The headline number only makes sense with the history attached. SambaNova last set a public valuation benchmark in April 2021, when a $676m Series D led by SoftBank Vision Fund 2 valued it above $5.1bn. Growth stalled after that: by December 2025, reports of Intel acquisition talks pegged SambaNova's value at roughly $1.6bn including debt — a steep markdown from its 2021 peak — before that deal fell apart in favour of a lighter-touch investment and collaboration arrangement. In February 2026, SambaNova unveiled its SN50 chip alongside a $350m Series E led by Vista Equity Partners and Intel Capital. Five months later, the $11bn Series F valuation represents roughly a 7x jump from the reported Intel offer price — one of the sharper valuation reversals we've tracked in AI infrastructure this year, and a reminder that headline valuations on private AI-infra companies are moving fast in both directions right now, not just up. We saw a comparable jump in public-market terms when Cerebras priced its IPO at a $56bn valuation in May, nearly doubling on its first trading day.

Dataflow versus GPUs: what an RDU chip is actually doing

Where GPUs bleed cycles on inference

A GPU is built around SIMT execution — thousands of CUDA cores running batches of largely independent operations, using high-bandwidth memory (HBM) as working memory. In practice, GPUs execute AI workloads kernel by kernel: each operation runs to completion, writes its result back to memory, and the next operation fetches it again before starting. That pattern is fine for training, where big batches keep compute units busy. It's a worse fit for the decode phase of LLM inference — generating one token at a time — where the hardware has to pull model weights from memory for every single token produced. On instruction-based architectures, that memory traffic, not raw arithmetic, ends up dominating latency.

What SambaNova's compiler does instead

SambaNova's RDU takes a different shape. Rather than a fixed instruction pipeline, the chip is a grid of Programmable Compute Units and Programmable Memory Units that a compiler configures per model — operations are laid out spatially, and tensors stream through them as a pipeline, the "dataflow" execution model the company's SN40L and newer SN50 chips are built around. Operator fusion keeps intermediate results on-chip instead of round-tripping to memory between every step, and SambaNova's published SN40L benchmarks claim over 85% HBM bandwidth utilisation from that approach. The chip pairs the design with a three-tier memory hierarchy — roughly 432 MiB of on-chip SRAM, 64 GiB of on-package HBM and up to 2 TiB of off-package DDR DRAM — aimed at fitting very large models on a single rack rather than spreading them across dozens of GPUs. The newer SN50, due to ship to customers later in 2026, is rated at 1.6 petaFLOPS at FP16 and 3.2 petaFLOPS at FP8, and can interconnect up to 256 accelerators over multi-terabyte-per-second links.

The inference-silicon landscape in one table

SambaNova isn't the only funded challenger trying to chip away at Nvidia's dominance of AI compute — Nvidia's market capitalisation sat around $5.1 trillion in mid-July 2026, with an estimated 70-80% share of the AI accelerator market, depending on whose count you trust. Here's how the four names builders actually weigh against each other line up right now.

SambaNova's named production customers back the "on-prem inference for regulated enterprise" positioning: JPMorgan Chase has selected SambaNova as an inference-infrastructure partner for secure, on-premises deployment of SN40 and SN50 systems, SoftBank Corp will be the first customer to deploy SN50 in its Japanese data centres, and Saudi Aramco is cited as a customer in SambaNova's own materials. That's a narrower go-to-market than Groq's or Cerebras's developer-facing cloud APIs — which matters for how quickly any of this reaches a smaller team's cloud bill.

What cheaper, credible non-Nvidia silicon means for builders renting compute

India — subsidised GPUs meet a new alternative

Most Indian AI teams aren't renting SambaNova racks directly — SambaNova has no announced India-specific cloud presence yet, and its customer list skews enterprise and on-premises. What this round changes is the backdrop those teams are renting against. India's compute story so far has been about subsidised GPU-hours and domestic capacity, from the IndiaAI Mission's GPU pricing programme to Neysa's $1.2bn Series B to build out India's own GPU cloud. A third well-funded, non-Nvidia inference architecture reaching enterprise scale strengthens the negotiating position of every buyer in that market, including Indian startups, because it gives cloud providers and inference platforms a genuine alternative supplier to point to when Nvidia allocation is tight. The same logic underpins why H100 rental prices falling toward $2/hour has been a 2026 story worth tracking — supply-side competition, wherever it comes from, tends to show up in the invoice eventually, even if it takes longer than the funding headline suggests.

UK — sovereign-compute ambitions get a second supplier

The UK's compute strategy has leaned heavily on a mix of sovereign infrastructure and hyperscaler partnerships — think Isambard-AI's frontier-safety experiments alongside private expansion such as Nebius committing £1.7bn to grow its London AI cloud hub. For UK teams building anything regulator-sensitive, a credible non-Nvidia inference vendor reaching enterprise production (SambaNova's JPMorgan Chase deployment is the clearest proof point so far) matters less as a direct rental option today and more as evidence that vendor diversity in AI inference silicon is becoming a real, fundable strategy rather than a slide in a pitch deck. That diversity flows through to the platforms UK builders actually touch: inference aggregators like Together AI, which raised $800m to scale open-source inference, exist precisely to abstract away which silicon sits underneath a given API call, and they get more room to negotiate on price the more credible alternatives to Nvidia exist upstream.

What to actually do with this, this quarter

• Don't switch inference vendors on a funding headline. Benchmark SambaNova, Groq, Cerebras and your current Nvidia-based setup against your own model, batch size and latency requirements before making any procurement decision.
• Ask your cloud or inference platform whether RDU or LPU capacity is on their roadmap. If you're already routing through an aggregator like Together AI or a regional GPU cloud, the underlying silicon diversity is their problem to manage — but it's worth asking, because it affects the prices you're offered later.
• Watch for the second close. SambaNova's Series F isn't fully done; a larger final round, or a fresh set of named investors, would be a further confirmation signal worth revisiting.
• Track all three challengers as a set, not in isolation. SambaNova, Groq and Cerebras collectively represent the first wave of non-Nvidia inference vendors reaching multi-billion-dollar scale in the same twelve-month window — that's a market-structure signal, even if no single one of them changes your architecture tomorrow.

So — does this change anything for you tomorrow?

Not immediately. SambaNova still doesn't have a self-serve retail cloud that most small Indian or UK teams can rent by the hour the way they can with Groq or Together AI — its named deployments are enterprise and on-premises, sold to buyers like JPMorgan Chase, SoftBank and Saudi Aramco who can commit to racks, not API credits. The pricing pressure this round represents will filter down to smaller builders over the next twelve to twenty-four months, through cloud and inference-platform competition, rather than showing up in anyone's invoice this week. What is worth registering today is the shape of the market: three well-funded, architecturally distinct, non-Nvidia inference vendors — a dataflow chip, a purpose-built LPU and a wafer-scale engine — have all reached multi-billion-dollar valuations within the same year. For builders in India and the UK who've spent the last two years assuming Nvidia allocation and pricing as a fixed cost of doing AI, that's the detail to keep watching, not the $11bn number itself.

Sources: General Atlantic, TechCrunch, Bloomberg and HPCwire. For more on the inference-silicon race, see our coverage of the Cerebras IPO.