Efficient Computer raised more than $97 million to move a chip out of the lab and into the data centre. The chip is the interesting part. Everything else is a bet on when that happens.
On 29 September 2026 the company, founded in 2023 and led by co-founder and CEO Dylan McDermott, said it had signed agreements for a Series B at a $650 million valuation with TQ Ventures leading. Total capital raised now stands at $173 million. The money goes into a 2 nm production run at GlobalFoundries, a facility in Texas, and a data centre chip that carries no name, no tape-out date and no shipping date. That is the whole story. A validated edge architecture is looking for a much larger market, and the search has a hardware dependency attached.
The edge product is shipping. The data centre product is a 2027 question.
The $650 million valuation prices the second one on the strength of the first. Investors are underwriting architecture transfer, not a data centre chip.
Post-money valuation, Series B
Pre-money is not disclosed. Agreements covered more than $97 million, bringing total funding to $173 million · Company release and Reuters, 2026
Dataflow, not a faster GPU
Efficient Computer builds spatial computers. The pitch is old and mostly discredited: put the data where the compute is, and stop paying to move it. What makes this version worth $650 million is the mechanism.
Fabric is a mesh of nodes on a 2.5D silicon interposer, packaged by Advanced Semiconductor Engineering. Each tile holds four compute engines and runs its own local memory. The design leans on an observation from 2021 research that the companies building NVIDIA's successors keep rediscovering: inference throughput and memory bandwidth scale badly together, so the industry keeps widening the bus and then discovering that power per query went up again. Fabric takes the opposite route. It lowers the memory wall by shrinking the distance data has to travel.
Electron E1 is the silicon proof point. Eight compute tiles. 640 mm² of interposer. Claimed up to 1 TOPS per watt at 8-bit. The company puts that at 10x to 100x the efficiency of current Intel and Nvidia architectures for real-time inference workloads at the edge.
We covered Positron's $875M round on 14 September, where the argument was that inference runs out of memory bandwidth before it runs out of arithmetic. Efficient Computer attacks the same wall from the other side, by moving less data rather than fetching it faster. Two architectures, one diagnosis.
McDermott founded the company in 2023 after building an internal accelerator at Tesla's Autopilot division and later leading AI silicon at Tesla's Dojo project. He was joined by Forrest Smith, who had run data centre software sales at Nvidia. The board includes former Intel CEO Pat Gelsinger, who chairs it. In February 2026 the company raised a $60 million round at a reported $420 million valuation, and Union Square Ventures and Eclipse Ventures were already on the cap table.
E1 is not a paper chip. It samples with more than 1,000 physical AI robots, the kind that walk, roll and grip. Efficient Computer says those systems run above real time on E1 today. Robot fleets are small, latency-sensitive and tolerant of fixed-function hardware. That is the beachhead.
What E1 proves, and what it does not
We have been running this on more than a thousand robots for over a year. No other architecture comes close on this workload.— Dylan McDermott, co-founder and CEO, Efficient Computer
The robotics claim is specific in a way chip marketing usually is not. A named customer class, a fleet count, a duration. It still leaves the number that matters unstated. How many tokens per second per watt, on which model, measured by whom. The 10x to 100x range runs against baselines the company selects, so it is not an independent result. Read it as a direction and a magnitude of ambition.
The absence of a data centre part is the more consequential gap. No name, no node, no tape-out. A 2 nm production run at GlobalFoundries buys one die, on one process, at volumes that suit edge devices. Rack-scale data centre economics depend on yields at volumes that only a committed product creates. The round funds the search. It does not fund the answer.
We think this can achieve more than 10x better energy consumption than other architectures for inference workloads.— Dylan McDermott, co-founder and CEO, Efficient Computer
More than 10x better energy consumption is a target, not a measurement, and it is the target that carried the round. Fewer joules per token decides the economics of inference at scale. That is the reasoning investors are being asked to accept, and it has one dependency worth stating plainly. Building for edge and building for a rack are different disciplines. Rugged, low-power, single-die designs meet commercial deployment schedules. HBM, chiplet packaging and thermal density define a different one, with fewer known customers inside a young company. Transfer is possible. It is not a given.
No customer names for the edge product. No revenue figure. No data centre chip name, process node, tape-out date or volume commitment.
The company does disclose team size: roughly 150 people, up from about 100 earlier in 2026.
The edge is a beachhead, the rack is the business
The capital is going somewhere specific. A 2 nm run at GlobalFoundries. A new facility in Texas. Team size up from about 100 to roughly 150. That is a company that has decided the architecture transfers and is spending on the transfer, with Series B money usually reserved for exactly that. Earlier stages paid for the silicon.
The valuation gap is the analytical hook. $650 million of post-money value now sits on $173 million of total capital raised. Investors are underwriting that gap with no disclosed revenue underneath it. Markets price gaps like this when the option looks large, and here the option is a data centre franchise attached to a product that already works somewhere else.
Against a base case that says no
The energy argument has a ceiling. The International Energy Agency put data centre electricity consumption at roughly 415 TWh in 2024, about 1.5% of global electricity, and projects roughly 945 TWh by 2030 under its base case. Its accelerated scenario reaches around 1,030 TWh by 2030. In both, the level is a rounding error against 2030 demand. Efficiency alone does not close the gap. Efficiency buys time.
Where the scenarios diverge is on the mix. Accelerated-server demand grows around 30% a year in the accelerated case against about 9% for conventional servers. Efficiency is the central variable there, because it changes how much capacity a constrained grid can host. Efficiency is worth a great deal in that world. It is not a substitute for new generation and transmission.
| Metric | 2024 | 2030 base case | 2030 accelerated |
|---|---|---|---|
| Data centre electricity | ~415 TWh | ~945 TWh | ~1,030 TWh |
| Share of global electricity | ~1.5% | ◐ ~3% | ◐ ~3.5% |
| Conventional server growth | — | ~9% a year | ◐ ~9% a year |
| Accelerated server growth | — | ◐ ~10% a year | ~30% a year |
IEA, Energy and AI, published April 2025. Figures are scenarios, not forecasts. The 2030 range between the two cases is under 10%.
That is the uncomfortable part of the bull case. A 10x efficiency gain on a 30% annual growth curve defers the constraint by a few years and changes the shape of the buildout. It does not remove it. Efficiencies of that magnitude have a habit of arriving as a premium product serving the workloads that can pay for them, which is a good business and a narrow answer to the grid problem.
The 10x to 100x efficiency range and the greater than 10x data centre target are Efficient Computer's own figures, measured on workloads of its choosing. No independent benchmark for Electron E1 exists.
The February 2026 round is reported at $60 million. Sources disagree on whether it was Series A or Series B, so the series label is left off here.
What the roundbook says
Reuters put the valuation at $650 million and the raise at more than $97 million, and named TQ Ventures as lead. Those are the verifiable numbers in this announcement. Everything above them, including the $97 million figure itself, is company-sourced. The phrase in the release is agreements, which is not the same as settled cash in every deal, though the distinction rarely matters at this stage.
The wider signal matters more than this round. Cerebras is pursuing large-scale datacenter compute on the same wager that general-purpose architectures are leaving performance on the table. Several challengers to Nvidia's inference position raised money in the same fortnight. Efficient Computer is the one claiming the answer already works, in a specific and narrow place, which is a stronger starting position and a narrower one.
Investors are being asked to underwrite a single technical transfer. If spatial computing carries a workload from a thousand small robots to a rack of thousands of accelerators, the multiple holds. If the data centre part slips a year, the same valuation is being asked to carry an edge business alone, and the comparable set changes completely. No customers, no revenue and no tape-out date mean that gap stays open until the company decides to close it.
What would have to be true for the data centre claim to land by 2027
Probability: 55% — the 2 nm GlobalFoundries run and the Texas facility are evidence of intent and of cash, and the Series B is explicitly sized for this part. Disclosure is cheap. Silence for four more quarters would be informative.
What supports the 2027 timeline
The Texas facility is a physical commitment with a construction schedule attached, not a lease.
Headcount rose from roughly 100 to roughly 150 in a year, which is the pace a tape-out team needs.
Confirmation criteria: a named data centre part, a process node, a tape-out date and one disclosed design win.
What argues against it
No disclosed revenue means no self-funded tape-out, so every schedule decision routes through the next raise.
Yields at data centre volumes are a different problem from yields at edge volumes.
Refutation criteria: no tape-out disclosure by mid-2027, or a Series C raised for working capital rather than tape-out.
A named data centre part with a process node and a tape-out date, announced separately from the funding.
First disclosed data centre customer, and whether it is a hyperscaler, a neocloud or an enterprise buyer.
Independent benchmarks on Electron E1, with the model, the baseline and the tokens per watt all published.
Hiring for HBM packaging and thermal roles, which would show the transfer in progress rather than intended.
Development scenarios
🟢 Architecture transfer works (25%)
Consequences: the $650 million mark looks cheap against a market where power per token sets the margin. Fabric becomes a second architecture rather than a curiosity, and the next round reprices the company as infrastructure rather than as an edge supplier.
🟡 Edge niche, data centre deferred (50%)
Consequences: a good business at a valuation set by edge comparables rather than infrastructure ones. The Series B did not destroy the company, it bought time that a lower entry price would have bought more of.
🔴 Efficiency claims fail contact with silicon (25%)
Consequences: the 10x to 100x range becomes the reason nobody underwrote the Series C, and $650 million becomes a down round instead of a floor. Spatial computing keeps its academic reputation.
Enjoying Nexi? Join the members-only tier for full access to our research and analysis.
Editor's note: the announcement of 29 September 2026 uses the word agreements for the Series B proceeds. Earlier reporting on the February 2026 round describes it at $60 million without agreeing on the series label. Efficiency figures throughout are company claims measured against company-selected baselines. No third-party benchmark for Electron E1 has been published as of writing.