Deep dive: Cheaper Tokens Without GPUs? Testing Tenstorrent's $6.2B Bet

Why today

This week, Tenstorrent said it had closed a $1.5B Series E at a $6.2B post-money valuation. That is more than twice the $2.6B mark from its Series D in late 2024. Trace Capital and Innovation X led the round. TechCrunch also names Fidelity and T. Rowe Price among the backers. Tenstorrent was founded in Toronto. It is run by the veteran chip designer Jim Keller and is now based in Santa Clara. The pitch is that its chips serve open-weight models and video models more cheaply than Nvidia GPUs. Earlier this year, Bloomberg reported takeover interest from Intel and Qualcomm. TechCrunch puts the reported offers at up to $5B. The company chose to stay independent and raise instead. So today's question is easy to ask and hard to answer. Once you count the software, how busy the hardware is and which models it can actually run, can non-GPU silicon really deliver cheaper tokens?

Plainly

Start with what cost per token actually means. It is the hourly cost of a machine divided by the number of tokens it produces in that hour. Everything else is detail, and the detail is where challengers win or lose. Here is a worked example using GPU numbers, because those are published openly. Spheron, a cloud provider that rents out Nvidia GPUs, prices an H100 at $3.38 per hour on demand. At 3,000 tokens per second, that node produces about 10.8 million tokens an hour. Divide the cost by the tokens and you get $2.50 per million tokens. Now suppose the node is busy only half the time. Utilization matters as much as the chip. A challenger chip attacks the cost side, meaning the price of the machine. Those machines are not like for like, but that price gap is the heart of the pitch. The token side is where it gets hard. Tokens per second depend on software: the compiler, the low-level kernels, and the serving layer that batches many users together. On Nvidia, that means CUDA plus open tools like vLLM, refined over many years by millions of developers. On Tenstorrent, it is an open-source stack called TT-Metalium, with a compiler called TT-Forge on top. Think of a cheaper, efficient engine that needs a custom gearbox for every car model. If each new model needs months of gearbox work, the fleet sits idle, and idle hardware is the most expensive kind. That is why model support matters. Open-weight models change constantly. Mixture-of-experts designs like DeepSeek send each token through a few specialist sub-networks, which puts new strain on memory and on the links between chips. A chip that runs last year's dense Llama well may struggle with this year's models.

What the evidence shows

What is established is mostly financial and commercial. TechCrunch and BetaKit report the round size, valuation and lead investors consistently. BetaKit reports that Tenstorrent has begun shipping Galaxy servers commercially and has about 1,000 employees. It also licenses its AI and RISC-V chip designs to LG and Samsung. Named customers and partners include the trading firm Virtu Financial and the data centre operator Equinix. Revenue is not established. The company would not tell BetaKit its total funding, or whether the round let existing holders sell shares. We found no revenue figure in any source. GPU.NET says licensing is currently the largest revenue source. If that is true, the cheaper-token story is not yet what pays the bills. The performance evidence is thinner still, because nearly every number comes from Tenstorrent itself. In 2024, EE Times saw a company demo of Llama 70B on a Wormhole box with eight chips. It served 32 users at 15 tokens per second each. The team called that result a work-in-progress. For its newer QuietBox 2 desktop system, Tenstorrent reports 476.5 tokens per second on Llama 3.1 70B. The biggest cost claim goes further. That is a 5x claim, and it is self-reported. Independent checks are scarce. In GPU.NET's words, "independently verified third-party benchmark coverage of Tenstorrent hardware is sparse". We found no Tenstorrent submission to MLPerf, the industry's shared inference benchmark. Spheron, which has its own stake in GPUs, adds a sharper caveat. Tenstorrent's own Galaxy figures of roughly 4,000 to 5,000 tokens per second come from runs "without a production serving layer". The Nvidia numbers they are compared against include real serving overhead. Spheron also writes that for mixture-of-experts models, "Expert routing does not fuse cleanly on the current Tensix mapping". That sits awkwardly beside the DeepSeek demo, and we cannot tell which account is more current. One more business signal: BetaKit reports a 7.5 percent workforce cut last year, during a shift from enterprise sales to developer sales.

How people read it

The same facts support several honest readings. The AGI-forward reading expects inference demand to outrun any single supplier. If models keep improving and agents run around the clock, the world needs every credible chip. On that view, a shipping rack at $110,000 has value before anyone audits a benchmark, and the step-up in valuation looks cheap. The AI-realist reading looks at history. Graphcore was valued at nearly $3 billion in 2020. Four years later it sold to SoftBank on undisclosed terms. Its CEO said the spending required was "an order of magnitude higher" than Graphcore could access. On this reading, good silicon is common. What is rare is a software ecosystem that keeps pace with new models, and Tenstorrent's revenue is still undisclosed. The market reading is about strategic value more than tokens. In December 2025, Nvidia agreed a deal valued at about $20B to license Groq's inference technology and hire its leaders. Large chipmakers clearly pay heavily for inference IP and teams. With Intel and Qualcomm reportedly interested, investors may be buying the option of a future acquisition as much as an operating business. The policy reading is about sovereignty. Tenstorrent sells to sovereign infrastructure providers and pitches an open stack that buyers can inspect and own. Toronto is still its largest office, with 200 staff, and some parts are built in Markham, Ontario. For a government that wants compute outside Nvidia's supply queue, an auditable alternative may be worth something even at a cost premium.

What it means for builders and investors

Our read, held provisionally, is this. The valuation looks priced on the open-weight inference thesis and on strategic interest, not on demonstrated unit economics. That is not a verdict on the engineering. It reflects what has actually been published. For builders, the opportunity is real if the numbers hold up. An inference company serving open-weight models on cheaper silicon could undercut GPU-based rivals on price. Spheron argues that the path from a supported model to a production endpoint takes months on Tenstorrent, against days to weeks on an H100 running vLLM. That engineering time is a cost, and it comes back with every new model family. For investors, the useful distinction is between the chip's cost per token and the company's cost per token. The second includes porting effort, idle capacity and the models customers actually want. When an inference startup claims a hardware edge, one question worth exploring in diligence might be: what is your measured cost per million tokens, on which model, at what utilization, and who measured it? A natural follow-up is how long it took to bring up the last new model. A specific answer with a date suggests a real advantage. A forecast suggests a hope. Independent data would change our view either way.

What to watch

Three things would settle much of this. The first is independent benchmarks. An MLPerf inference submission would turn claims into evidence. So would a third-party test of Galaxy Blackhole on a current mixture-of-experts model with a real serving layer. The second is public prices. If an inference platform lists per-token prices for open models served on Tenstorrent, the market has run the benchmark for us. As of mid-2026, GPU.NET reported that Galaxy hardware was not yet available on public cloud marketplaces. The third is revenue. Keller has said an IPO is the target, and a filing would show revenue and how it splits between hardware and licensing. None of these has a firm date. Until one arrives, cheaper tokens on Tenstorrent remain a claim made by Tenstorrent.