Seven months separate a rescue sale from a sevenfold markup
Last December, Bloomberg reported that Intel was closing in on a deal to buy SambaNova for about 1.6 billion dollars, debt included. The Palo Alto chipmaker, founded in 2017 by Stanford professors and valued at 5 billion dollars back in 2021, looked like another casualty of Nvidia's grip on AI hardware. The talks collapsed. In February the company took a 350 million dollar Series E and unveiled its SN50 chip instead. On July 8 it announced a 1 billion dollar first close of its Series F at an 11 billion dollar post-money valuation, led by General Atlantic.
The roster behind the round reads like a pre-IPO order book. T. Rowe Price, Capital Group, Seligman Ventures, BlackRock funds, Vista Equity Partners, the Qatar Investment Authority and Intel Capital all participated, and chief executive Rodrigo Liang says a second close should wrap within weeks. Liang is specific about where the money goes: locking down the supply chain so the company can deliver hardware against orders over the next twelve months. For a company that was shopping itself in December, that is a violent change of fortune.
Inference is where your AI spending actually lives
Training a model is a one-time project. Serving it is a bill that arrives every month, every time an employee drafts a document or a customer hits a chatbot. That recurring line, called inference, is where enterprise AI budgets are migrating, and it is exactly the market SambaNova's reconfigurable dataflow chips target. The company serves open-weight models such as DeepSeek and Llama at high speed, in the cloud or inside a customer's own data center, and counts Saudi Aramco and SoftBank among its customers.
The jump from 1.6 to 11 billion dollars in seven months is a bet on that migration. For an operator, the practical reading is blunt: per-token and per-query costs compound quietly across every workflow you automate. The unit price of inference deserves the same scrutiny you give freight rates or card processing fees, because within two years it will be a comparable line item on your accounts. Finance teams that already track cloud spend line by line should give inference its own budget code this year.
JPMorgan's on-premises order is the loudest signal in the deal
Alongside the funding, JPMorgan Chase named SambaNova an inference infrastructure partner. The bank will run SN40L and SN50 systems inside its own facilities to power secure inference on enterprise workloads, and its engineering leadership publicly flagged the speed and security profile of the architecture as the draw. SambaNova says demand now comes from three buckets: sovereign clouds, specialist AI clouds and large enterprises.
When one of the most heavily regulated banks on earth validates a non-Nvidia stack for in-house AI, procurement teams everywhere gain cover to evaluate the same route. If your business handles client data under GDPR, health rules or financial supervision, on-premises inference just moved from an exotic experiment to a legitimate line in your next infrastructure review. Expect midsize software vendors to follow the bank with on-premises options of their own.
A funded second source hands you negotiating leverage
Most AI vendors you buy from resell compute that ultimately runs on Nvidia hardware, and their pricing reflects that single choke point. Every credible alternative that banks a billion dollars of runway weakens take-it-or-leave-it pricing across the stack. Ask your provider which hardware sits behind the API you pay for, and ask how their unit costs have moved over the past two quarters. Silence is an answer too.
SambaNova's SN50 begins shipping to customers in the second half of 2026, and the company co-develops products with Intel, whose chief executive Lip-Bu Tan has chaired SambaNova's board. Hardware competition at this scale pushes unit costs in one direction. Signing a multi-year inference contract at today's rates means paying 2026 prices for 2028 compute.
Three checks to run before your next contract renewal
First, re-quote. Benchmark per-million-token pricing across at least three providers every quarter and put the numbers in front of your incumbent. Second, stay portable. Favor open-weight models where quality allows, so a workload can move between chips and clouds without a rewrite, and keep your prompts and evaluation data in a form you own.
Third, if you operate under regulatory supervision, price an on-premises or sovereign inference option against your current cloud bill. Then watch two dates: SambaNova's second close in the coming weeks, and SN50 deliveries in the second half of 2026. If both land on schedule, list prices for fast inference will move again, and your renewal should be timed to capture it.
Read next: Bizay raises 48.75 million euros at home and aims the money at America | Paradigm's $1.2 Billion Fund Takes Crypto Money Into AI and Robotics



