Nvidia’s AI Infrastructure Gains Depth: Vera Benchmarks Arrive as Blackwell Deals Go Live

Nvidia's Vera CPU outperforms x86 in benchmarks; IREN orders $1.6B in Blackwell systems, signaling Nvidia's full-stack AI shift. Revenue hit $81.6B.

Nvidia’s strategy is no longer just about selling the fastest chips. The company is now building a full-stack AI empire, and two separate developments this week underscore how far that ambition reaches: a new in-house CPU that can hold its own against x86 incumbents, and a concrete Blackwell cloud order that turns a multi-billion-dollar partnership into hardware on the ground.

The Vera processor, tested by the technology site Phoronix on May 26 at Nvidia’s Santa Clara headquarters, marks the company’s most serious push yet into server CPUs. Stuffed with 88 custom Olympus cores, the chip delivered up to 1.2 terabytes per second of memory bandwidth through an LPDDR5X subsystem, maintaining 90 percent of that theoretical peak in the STREAM TRIAD benchmark. Per core, Vera achieves more than four times the memory throughput of conventional x86 processors. Compared with its predecessor Grace, total performance improved by a factor of 1.6; against a current 128-core x86 chip, Vera claims a 50 percent lead. The sort of raw compute needed to compile a standard Linux kernel took a single Vera socket just 20 seconds — the fastest result in the entire Phoronix test suite.

Those numbers come with obvious caveats. The tests ran on a pre-production unit using workloads selected by Nvidia itself. Power consumption and clock-frequency monitoring were disabled because the processor’s energy management hasn’t been fully integrated into the Linux kernel yet. The early signal is strong, but a complete picture — including performance per watt and comparisons with rival server platforms — will have to wait until production systems arrive, likely sometime in 2026.

Meanwhile, the Blackwell ecosystem is moving from paper to data centre floors. On Tuesday, the cloud and crypto-mining firm IREN announced a roughly $1.6 billion purchase agreement with Dell for air-cooled Blackwell systems. The hardware — including GPUs, servers, storage, networking gear, integration and warranties — will be installed at IREN’s existing site in Childress, Texas, with initial operations targeted for early 2027.

The order builds on a five-year AI infrastructure cloud service contract that IREN signed with Nvidia on May 7, valued at about $3.4 billion. Under that deal, IREN provides managed GPU cloud services for Nvidia’s internal AI and research workloads. The Dell purchase is not a new Nvidia contract but rather the procurement that turns a known plan into physical capacity. The two companies also announced a strategic partnership that could ultimately support up to 5 gigawatts of DSX-near AI infrastructure. As part of the arrangement, Nvidia received the right to buy as many as 30 million IREN shares at $70 each — a potential $2.1 billion equity stake.

Both developments arrive against a backdrop of blistering financial performance. In the first quarter of fiscal 2027, Nvidia posted record revenue of $81.6 billion, up 85 percent year over year. The data-centre segment alone grew 92 percent to $75.2 billion. Under the company’s revised reporting structure, which splits data centre into Hyperscale and ACIE (AI Clouds, Industry and Enterprise), IREN’s cloud project fits squarely into the ACIE category — a segment that will become more visible in future earnings calls. For the current quarter, Nvidia expects revenue of roughly $91 billion, plus or minus 2 percent, a figure that explicitly excludes any data-centre compute sales from China.

Shareholders are also seeing cash come back. Nvidia returned about $20 billion through buybacks and dividends in the first quarter. In May the board authorised an additional $80 billion repurchase programme, and the quarterly dividend was hiked from $0.01 to $0.25 per share, payable on June 26.

The stock closed Tuesday at €184.46 in European trading, roughly 8 percent below its 52-week high of €201.05. It remains up more than 54 percent over the past 12 months, although the week-to-week move is a 4 percent decline. At a market capitalisation of about $5.2 trillion, exceptional execution is already priced in. That makes technical milestones like the Vera benchmark increasingly important: they reinforce the thesis that Nvidia is not merely dominating the GPU market but gaining control of the entire AI data-centre stack — from processors to networking to software.

For IREN, the Blackwell ramp is expected to lift its annualised revenue run rate from $3.7 billion to $4.4 billion, a figure that includes anticipated revenue from its Microsoft contract, the Nvidia cloud deal and planned GPU expansion projects. None of that is fully contracted, so actual results may vary. That uncertainty, along with the absence of Chinese data-centre revenue from Nvidia’s guidance, sets up the next test for the company: supply-chain execution, Blackwell’s production pace and the speed at which partners like IREN can bring live compute capacity online.