Nvidia-Backed doubleAI Spent About Half Its $220M on B300s

doubleAI, a Tel Aviv startup backed by investors including Nvidia, is set to debut Maxwell, a model it says lowers companies’ computing costs, The Wall Street Journal reported on 19 October 2026. The startup has raised $220 million in total and used roughly half of that funding to buy 1,000 Nvidia B300 chips for its own needs, the Journal reports.

As we read it, the company is on both sides of the compute shortage. It plans to charge customers a share of what it saves them on compute, and it bought its own chips rather than rely on renting them. Chief executive Amnon Shashua told the Journal: “You cannot rely just on renting compute.”

Business Pill · GPUS

A short explainer on why AI runs on GPUs. It teaches the general idea only and says nothing about any company in this story.

The Launch

doubleAI describes Maxwell as the first of what it calls “code execution models,” or CEMs, which it says put the AI “between code and silicon,” the Journal reports. Unlike language models that help people write code, CEMs live inside running programs and operate without human involvement, according to the report.

On its own Maxwell site, doubleAI says the model rewrites code while it runs, for the exact hardware and data in use, and that “Every improvement is formally proven equivalent to the original code.” The source code itself stays unchanged, the company says. It runs on both CPUs and GPUs.

The big challenge was verifying the code, Shashua told the Journal, and doubleAI found a way to perform that verification formally and mathematically, he added. Maxwell is due to become more widely available later this month, the Journal reports; the company’s site offers early access.

What Was Posted on X

doubleAI chief executive Amnon Shashua, 4 August 2026: he says doubleAI rewrote pandas’ kernels for modern hardware as a drop-in replacement that runs 38.4x faster on the PDS-H benchmark. It is the company’s own claim.

Shashua, 5 October 2026, asking whether AI can reach superoptimisation at internet scale, with every workload tailored to its exact hardware, and announcing a talk on it by doubleAI’s Gal Beniamini at Nvidia’s GTC in Berlin on 22 October. As we read it, that is the pitch the Journal describes for Maxwell, made two weeks before the launch.

How It Plans to Charge

doubleAI plans to charge customers based on how much they save on compute, the Journal reports. In Shashua’s example, if a customer’s compute bill falls from $10 million to $6 million, doubleAI keeps some percentage of the $4 million saved. He did not say how much.

The startup said it cut one early customer’s compute costs by 67%, according to the Journal. LTX, which builds open-weight multimodal models, found that Maxwell sped up its model inference twofold. “The speedup directly translates to cost,” LTX co-founder and chief executive Zeev Farbman told the Journal.

Shashua expects companies with massive compute bills for training and inference, such as AI labs like OpenAI and Anthropic, to see the most benefit, the Journal reports. For the major AI labs, he said, even a 10% compute saving could represent hundreds of millions of dollars.

Buying Its Own Chips

To meet its own computing needs, doubleAI used roughly half of its funding to purchase 1,000 Nvidia B300 chips, the Journal reports. Shashua said the startup did not want to scramble for GPUs. “Prices are going up,” he said.

The Journal cites SemiAnalysis data showing the hourly cost of renting Nvidia H100 chips on a one-year contract has risen 60% in the past year. Our arithmetic: roughly half of $220 million is about $110 million, or roughly $110,000 per chip across 1,000 B300s. The Journal does not break that figure down.

“More hardware is being built, but the need for compute is rising faster than the ability to build GPUs and CPUs,” Shashua told the Journal. For a cost comparison at the chip level, see our report on Google’s TPUv7 costs.

What doubleAI Says Maxwell Does

doubleAI’s Maxwell site lists its own results: 38× faster data processing pipelines and 4.25× faster AI inference kernels. For LTX-2.5, it says 4K generation took 134 seconds per clip on two B200s, down from 250 seconds, or 1.87× faster.

The site also shows two demos. A Whisper speech-to-text run on an RTX 4090 took 49.82 seconds with Maxwell against 106.23 seconds without, and a NetworkX graph job took 17.50 seconds against 61.44. Our arithmetic puts those at about 2.1× and 3.5×. A research post on the site reports a cuGraph version 3.6× faster on average.

These are the company’s own benchmarks. We have not tested them, and the Journal reports the 67% saving as doubleAI’s statement.

Who Backs It

The Journal names Lightspeed Venture Partners, Bessemer Venture Partners, Nvidia and Dell among doubleAI’s investors. The company’s own homepage lists NVIDIA, Lightspeed, Bessemer Venture Partners, Dell Technologies Capital, Pitango, Greenfield Partners and BRM Group under “Backed by.”

doubleAI is two years old and has about 40 employees, the Journal reports. Shashua, who co-founded it, also helped launch Mobileye and is still its chief executive; in July Mobileye said he would step down once it names a successor.

The Structural Read

As we read it, the pricing ties doubleAI’s revenue to its customers’ compute bills: in the Journal’s example, a bill that falls from $10 million to $6 million leaves $4 million of savings, of which doubleAI keeps an undisclosed percentage.

As we read it, the chip purchase is the startup’s own answer to rising rental prices: the Journal cites SemiAnalysis data showing one-year H100 rental costs up 60% in a year, and Shashua said “Prices are going up.”

As we read it, the evidence so far is the company’s: the 67% saving is doubleAI’s statement, the speedups are its own benchmarks, and the one named customer, LTX, reports a twofold inference speedup.

Three Implications

PRICING ON OUTCOMES The Journal reports doubleAI will charge based on how much customers save on compute, not a flat fee; the percentage is not disclosed.

STARTUPS BUYING THEIR OWN CHIPS The Journal reports doubleAI used roughly half of its funding on 1,000 Nvidia B300s; our arithmetic puts that at roughly $110,000 per chip.

WHERE THE SAVINGS MATTER Shashua told the Journal that for the major AI labs even a 10% compute saving could represent hundreds of millions of dollars.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework puts it this way: “Inference accounts for 80 to 90 percent of the lifetime cost of a production AI system.”

As we read it, that is the cost line Maxwell is aimed at: doubleAI says it speeds up AI inference kernels and the Journal reports LTX saw a twofold inference speedup. The framework was published in March 2026 and does not discuss doubleAI.

What Is Not Established

doubleAI did not say what share of savings it will keep. The 67% saving and the speedups are the company’s figures; the early customer is not named. The Journal does not say whether the chip spending covers servers or hosting, or when the $220 million was raised. The per-chip figure is our arithmetic on a “roughly half” estimate. doubleAI had not posted about the launch on its X account when we checked on 19 October.

The Bottom Line

doubleAI, backed by investors including Nvidia, has raised $220 million and used roughly half of it to buy 1,000 Nvidia B300 chips, the Journal reports, as it launches Maxwell. It plans to charge a share of the compute costs it saves customers, and said it cut one early customer’s costs by 67%.