Summary: TechCrunch broke the news on July 8, 2026: SambaNova Systems, an AI chip maker, has banked the first $1 billion of its Series F at an $11 billion valuation. A layered memory stack — static random-access memory (SRAM), high-bandwidth memory (HBM), and large-capacity memory — lets models hot-swap within milliseconds, and caching of input tokens pulls Time to First Token down. General Atlantic led it, with returning backers such as Intel chipping in again — just five months after the company’s previous $350 million Series E round.
TechCrunch broke the news on July 8, 2026: SambaNova Systems, an AI chip maker, has banked the first $1 billion of its Series F at an $11 billion valuation. General Atlantic led it, with returning backers such as Intel chipping in again — just five months after the company’s previous $350 million Series E.
Five Months On, Another Billion-Dollar Round
The $1 billion is only the first close. CEO Rodrigo Liang expects a second close, with more investors attached, to wrap up within weeks. Beyond lead backer General Atlantic sit Intel — in since Series C — plus T. Rowe Price, BlackRock, Capital Group, Battery Ventures, Vista Equity Partners, and the Qatar Investment Authority. What the money is for is spelled out plainly: scale the business and lock down the supply chain over the coming 12 months.
The cadence tells its own story. February 2026 brought the $350 million Series E, with Vista leading, and the next-generation SN50 chip debuted at the same moment. Five months on, another $1 billion has landed at an $11 billion valuation — proof the capital race in inference compute keeps speeding up, and that investors will pay above the odds for a genuine alternative to the GPU establishment.
Meet the SN50: A Fifth-Generation RDU Tuned for Agentic Inference
The company’s core product is the RDU (Reconfigurable Dataflow Unit), a dataflow architecture that handles inference workloads off the GPU-cluster path. Unveiled on February 24, 2026, the fifth-generation SN50 fits 16 chips inside each SambaRack system. By the company’s figures, on agentic inference benchmarks run with Llama 3.3 70B, it posts 5X maximum speed and better-than-3X throughput against NVIDIA’s Blackwell B200; stretched across 256 accelerators, it carries models up to 10 trillion parameters with context windows reaching 10 million tokens.
Design-wise, the SN50 treats inference as a problem of moving data. A layered memory stack — SRAM, HBM, and large-capacity memory — lets models hot-swap within milliseconds, and caching of input tokens pulls Time to First Token down. That suits agentic loads: agents hop between models constantly, and on conventional clusters each hop means reloading weights — a costly latency tax. A rack draws roughly 20 kW on average and needs only air cooling, easing retrofit costs in existing data centers. Shipments begin in the second half of 2026, with SoftBank the first deployment partner.
Who Buys: SoftBank, JPMorgan, and Beyond
The round landed alongside a JPMorganChase announcement naming SambaNova its inference-infrastructure partner; the bank will run security-sensitive inference workloads on SN40L and SN50 systems within its own facilities. Liang reads the deal as a message to banking at large: institutions that lean hard on AI are not confined to the public cloud.
Saudi Aramco and several Japanese companies round out the disclosed customer list. SambaNova courts sovereign clouds, neoclouds, and large enterprises — the customers with the sharpest demands around data sovereignty and on-premises deployment. The earlier SN40L sells in cloud and on-premises form alike, while the SN50 leads with frontier models on one rack.
Intel: Investor, Almost-Acquirer, Now Partner
Intel’s position is the trickiest part of the story: an investor since Series C, it came back for this round too. Back in December 2025, Bloomberg reported that a roughly $1.6 billion acquisition of SambaNova by Intel had been close before the negotiations lost momentum. By February 2026 the two had settled on a multi-year collaboration instead: co-developing and jointly selling AI inference solutions on Intel’s Xeon platform.
For Intel, pairing an equity stake with joint sales tethers a fast-growing inference chip company without an acquisition; for SambaNova, it preserves independence and bargaining power. Asked whether the company will stay independent, Liang declines to commit, saying only that momentum points to an eventual IPO. At $11 billion, the entry price for any acquirer towers over where it stood five months ago — the buyout window is shutting fast.
Where the Inference Market Goes From Here
SambaNova plays in premium inference — multi-trillion-parameter frontier models at single-rack scale, head-to-head with NVIDIA’s GPU ecosystem. Cerebras and Groq run the same track, betting that agentic AI will detonate inference demand: once AI systems call models in sequence and run for long stretches, inference cost and latency — not training performance — become what buyers weigh first.
For purchasers, two takeaways matter. One, the inference-hardware menu keeps expanding; GPUs are no longer the sole route, and every credible new entrant enlarges the room to negotiate on cost and latency. Two, once shipments begin in the second half of 2026 the SN50’s real-world numbers get tested; whether the claimed multiples survive that scrutiny decides if this $1 billion was a bubble or a ticket. Either way, a rack-scale inference specialist raising at $11 billion says where the market believes the next phase of AI infrastructure spending belongs.