$11 billion. That is the valuation SambaNova just reached, roughly seven times what Intel was reportedly ready to pay for the entire company in December. The acquisition talks collapsed. The startup raised $1 billion on its own terms instead.
JPMorganChase picked it as an inference-infrastructure partner, deploying SN40L and SN50 systems for AI inference inside the bank's own firewalls.
Intel moved from would-be buyer to backer. The reported $1.6 billion takeover stalled; it invested instead, and its CEO still chairs the board.
SambaNova builds custom chips and full systems for AI inference, the work of running trained models, not training them. The company's argument is that enterprises in regulated industries want the biggest models executing on their own hardware rather than in someone else's cloud. JPMorgan is the reference case for that bet.
TIMELINE: SambaNova — from $1.6B takeover talks to $11B
─────────────────────────────────────────────────────────────
2021 ──── Dec 2025 ──── Feb 2026 ──── Jul 2026 ──── 2027
🚀 ⚠️ 🤝 ◉ NOW 🔥 NEXT
Series D Intel offer Series E Series F IPO?
$5.1B $1.6B $350M $1B planned
Funding and valuation trajectory from company announcements, TechCrunch, Bloomberg and CNBC, 2026.
The RDU bet, explained in one architecture
The heart of SambaNova's approach is the Reconfigurable Dataflow Unit (RDU), a chip built to keep data flowing across a programmable grid instead of shuttling it through a fixed processor core. Each generation reconfigures that grid in software. The fifth-generation SN50, unveiled in February 2026, is the version now aimed at agentic inference workloads.
The numbers it publishes are pointed. The SN50 claims up to 5x the maximum speed of competitive chips and 3x the throughput of Nvidia's B200 on agentic inference tasks, and it can scale to 256 chips across racks to run models up to 10 trillion parameters. The Register's benchmark coverage described the systems breathing new life into aging Nvidia GPUs: the SN40 and SN50 handle the decode portion of inference five to ten times faster, freeing existing GPUs for other work.
None of that replaces Nvidia. It widens the second lane.
How a $1.6 billion offer became an $11 billion round
In December 2025, Bloomberg reported that Intel was in advanced talks to acquire it for roughly $1.6 billion including debt. The deal was close enough that the price leaked and the market priced it in.
The talks stalled. In January 2026, Bloomberg reported SambaNova was seeking up to $500 million in new funding instead. In February, the two companies announced a different arrangement: Intel would invest, and the pair would co-develop AI inference systems built on Intel Xeon servers. The investment received US antitrust clearance in May, and a planned additional $15 million would take its stake to roughly 9%.
The arc matters more than the destination. Seven months after a takeover at $1.6 billion, the same company raised $1 billion at $11 billion. Intel's CEO, Lip-Bu Tan, has chaired the board since 2017 and recused himself from the collaboration discussions. Intel Capital participated in the new round anyway.
JPMorgan and the on-premises thesis
The commercial signal in this round is a customer, not the cheque. JPMorganChase selected it as an inference-infrastructure partner, deploying SN40L and SN50 systems for secure, on-premises AI inference behind the bank's own firewalls. SambaNova announced the relationship; JPMorgan has not issued an independent confirmation.
Darrin Alves, chief information officer of infrastructure platforms at JPMorganChase, said the bank's AI infrastructure must meet a high bar for performance, control and reliability. CEO Rodrigo Liang reads that as a broader market statement. Regulated industries want private inference under their own control, he argues, and the deal signals that the on-premises thesis is now bank-grade.
Earlier customers fill out the same picture. SoftBank was the first SN50 customer, folding the chip into sovereign AI data centers in Japan. The company cites deployments across providers in Australia, Europe and the UK, and counts Hugging Face and Meta among its users.
Why inference is where the spend is going
The financing round is large by any measure, but it sits inside a bigger rotation. Research firm Grand View values the AI inference market at $97.24 billion in 2024, heading to $253.75 billion by 2030. The centre of gravity is moving from training models to running them at production scale.
That shift is already visible in the competitive field. Cerebras, a rival inference-silicon maker, priced an IPO in May 2026 and raised $5.55 billion. Nvidia agreed to a roughly $20 billion non-exclusive licensing and asset deal with inference startup Groq in December 2025, keeping Groq independent while locking in its technology. Three different capital structures, one bet: the inference layer is worth owning.
Turning points
Four decisions shaped the outcome. First, the Intel takeover talks stalled in late 2025, freeing it to stay independent. Second, the two companies converted that failure into a co-development partnership in February 2026, giving it access to Intel's scale without surrendering control. Third, the SN50 launch in the same month gave the company a concrete product story to raise against. Fourth, the JPMorgan selection turned that story into a reference sale.
Each step looks obvious in hindsight. None of them was obvious at the time.
What the round means for investors
The first-close structure matters. A $1 billion first close leaves room for a second close with more investors, so the final total can land higher. The cap table already mixes growth equity, mutual funds and sovereign capital: General Atlantic led, with Seligman Ventures, T. Rowe Price, Capital Group, BlackRock, the Qatar Investment Authority and others in the syndicate.
Liang told CNBC the company is strongly considering an IPO in 2027, most likely in the US. Total funding now exceeds $2 billion. A public listing would make it one of the first pure-play inference-chip challengers to face the public market, a test investors in Cerebras have already begun to price.
The risks are symmetrical with the thesis. The JPMorgan relationship is announced, not jointly confirmed, and its speed claims rely on its own benchmarks until independent labs weigh in. Nvidia holds roughly 80% of the data center accelerator market and has both the silicon cadence and the balance sheet to answer an on-premises challenger. A second close that fails to materialize would also read poorly, since the round was deliberately staged.
Keep the architecture simple. The wager is that enterprises will pay a premium to run AI where their data lives. JPMorgan is the first bank-sized proof point. The second one will be the tell.