Weekly: On remote access compute
14 min read.
Highlights
Remote access compute. Michelle Nie, a visiting fellow at CNAS, published a piece calling for the closing of the remote-access compute loophole of the chip sanctions. As I frequently say here, it is a widely known and widely abused loophole and many Chinese AI models are using cutting-edge Nvidia chips without having to purchase/smuggle them directly. One recommendation is a Know Your Customer (KYC) requirement for cloud compute providers. Congress is looking into this loophole through the Remote Access Security Act, which has passed the House and is favoured to pass in the Senate.
Pooling data centres through some remote access is one way for Europe to overcome compute and energy bottleneck, say researchers at RAND. Networking compute across different data centres across great distances is difficult but the technology has gotten better in recent years.
Long-term contracts. Asa Fitch of the WSJ writes about long-term contracts which have become increasingly popular. For instance, Micron has signed many long-term agreements (LTAs), up to five years, with its customers in recent months. The main pro of LTAs is that it gives more visibility for Micron and peers in what is famously a volatile and cyclical sector. The con, however, as Fitch explains, is that it could exaggerate volatility in a downturn with unused chips piling up in inventory. I wrote about this dynamic earlier this week.
Nvidia’s latest hardware. WIRED and SemiAnalysis both offer breakdowns of Nvidia’s latest AI server, Rubin GPUs, and Vera CPUs. WIRED does a good job for the general and educated audience, and SemiAnalysis is of course the gold standard of technical, engineering analysis. WIRED makes the astute overarching claim, “Nvidia wants to own every chip inside AI data centers.”
Thanks for reading.
Table of Contents
Rafael Andersson Lipcsey and Maximilian Negele, “Pooling Europe’s compute: The promise of distributed training for European frontier AI,” RAND, 07/16/2026.
Michelle Nie, “Closing the Remote Access Loop,” Issues in Science and Technology, 07/23/2026.
Asa Fitch, “The Massive Supply Deals Feeding the AI Frenzy Are No Sure Thing,” WSJ, 07/20/2026.
Lauren Goode, “Nvidia Wants to Own Every Chip Inside AI Data Centers,” WIRED, 07/21/2026.
Alec Ibarra, Bryan Shan, Daniel Nishball, et al., “Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis,” SemiAnalysis, 07/23/2026.
Joel Khalili, “This Former Intel CEO Wants to Jumpstart Moore’s Law With Light,” WIRED, 07/21/2026.
The Economist, “Does SpaceX’s plan for a data centre in space add up?,” The Economist, 07/22/2026.
1.
Rafael Andersson Lipcsey and Maximilian Negele, “Pooling Europe’s compute: The promise of distributed training for European frontier AI,” RAND, 07/16/2026.
Europe lags significantly behind the United States and China in frontier AI capability. The gap has multiple underlying causes, with a lack of European capital chief among them. This report, however, focuses on a more specific challenge: to train the most advanced AI models, compute has until recently had to be concentrated at a single site, a condition particularly hard to meet in Europe given three bottlenecks: power (slow infrastructure buildout and insufficient power at any single site), compute (a small and fragmented compute stock) and politics (the difficulty of political coordination across member states).
Distributed training, a set of techniques allowing geographically separated datacentres to function as a single training resource, has matured rapidly since 2023. This report assesses its relevance to European frontier AI capability, drawing on a literature review, expert consultations and quantitative estimation of compute and power requirements. The analysis finds that distributed training offers relief on the power and political-coordination bottlenecks. As for compute, it cannot solve the lack of chips itself, but it changes whether Europe’s existing and planned compute can support frontier training at all: without distributed training, the same chips remain fragmented across clusters individually too small for frontier work. The report sets out policy recommendations in two areas: preparing European physical infrastructure for distributed training, and simplifying and harmonising cross-border regulation.
Distributed training offers substantial relief on Europe’s power and political-coordination bottlenecks but does not narrow Europe’s underlying chip gap. What it does change for compute is whether the chips Europe has can be pooled into a single resource for frontier training, rather than remaining fragmented across individual clusters too small for frontier work.
Utilities are typically more willing to commit power to a portfolio of smaller sites than to a single multi-gigawatt facility, and connection requests for smaller sites can proceed in parallel through different national queues, compressing overall timelines. In addition, the demand flexibility that distributed training enables in datacentres can free up additional grid capacity.
Distributed training does not add chips. As of the beginning of 2026, European operational capacity stands at approximately 123,000 H100-equivalents, against roughly 1.4 million in the United States. Including all confirmed and likely planned capacity, Europe reaches around 3.2 million by 2030, while the United States reaches at least 19.4 million. What distributed training changes is whether this stock can be pooled into a single resource large enough for frontier training.
2.
Michelle Nie, “Closing the Remote Access Loop,” Issues in Science and Technology, 07/23/2026.
As Asad Ramzanali argues in “Why the Cloud Needs Competition” (Issues, Winter 2026), cloud computing has become critical infrastructure for everyday American life and national security. It is also the primary platform on which frontier artificial intelligence, the leading edge of AI capability, is built and trained—and therefore central to US-China competition and securing American AI leadership. It’s time for the US government to govern it accordingly by addressing a glaring gap: the lack of appropriate oversight into operations on the cloud relevant to national security.
Chinese entities have exploited this gap, quietly accessing American cloud infrastructure to run AI workloads on chips they are banned from purchasing directly, undermining the export controls that represent the United States’ greatest lever in maintaining its leadership in AI. One case involving Chinese government-linked Shenzhen University, which remotely accessed Nvidia A100 and H100 chips via Amazon Web Services, was discovered only when Reuters journalists combed through Chinese procurement documents.
It’s time for the US government to govern it accordingly by addressing a glaring gap: the lack of appropriate oversight into operations on the cloud relevant to national security.
A know-your-customer (KYC) regime, as Ramzanali proposes, would address the lack of oversight into commercial cloud activity. Requiring cloud providers to verify who is accessing their infrastructure and report that information to federal authorities would close the gap that allowed Chinese entities to access restricted compute undetected. Such a regime has strong bipartisan support, having already been proposed by both the Trump and Biden administrations.
Congress is making progress on closing the remote access loop. The Remote Access Security Act, which broadens compute export controls to include remote access through the cloud, recently passed the House and is under active consideration by the Senate. However, success requires implementation of an effective KYC regime. If the bill passes, cloud providers would have to ensure that they’re not providing services to entities of concern—and a strong KYC program is the starting point for fulfilling the requirement.
3.
Asa Fitch, “The Massive Supply Deals Feeding the AI Frenzy Are No Sure Thing,” WSJ, 07/20/2026.
Contracts worth billions of dollars have become the commercial glue holding the AI boom together. But investors shouldn’t rely on them sticking if the boom fades.
Contracts to supply computing power for artificial-intelligence calculations have become such a fixture of the craze that whole industries have begun reorganizing around them. AI suppliers say these arrangements give them unprecedented visibility into their future revenue, allowing them to wow investors with promises of bumper sales and profits ahead.
The computer-memory business may be the most extreme example. Memory suppliers and their customers have sought out longer-dated deals in recent months.
This is because of the explosive growth of autonomous AI agents, which are memory-intensive. And that is turning what historically has been a cutthroat, cyclical business where price competition is fierce into something much more stable.
Micron has been an especially active dealmaker. Its “strategic customer agreements” typically last five years and are take-or-pay. This means the buyer has to pony up whether they take shipment of the memory or not. Last month, Micron Chief Executive Sanjay Mehrotra said on an earnings call that these agreements would supply more than half of the company’s revenue in the years ahead.
But while the long-term contracts are helping drive the upswing, there is reason to wonder whether they provide any certainty in a downturn.
Indeed, if demand for memory should ebb before supply contracts expire, those will likely be renegotiated or extended. Chip suppliers don’t like sending products to customers who aren’t going to use them.
If they do, the chips will sit on the shelf until demand returns. And at that point, customers will draw down inventories before they buy new chips, delaying revenue for the chip makers.
Suppliers also don’t like to force products on customers who don’t want them, especially if competitors are being more flexible. Pushing unwanted memory into the supply chain could fray long-term customer relationships.
You don’t have to dig too deep into tech history to find evidence suggesting that suppliers will be accommodating in a downturn.
4.
Lauren Goode, “Nvidia Wants to Own Every Chip Inside AI Data Centers,” WIRED, 07/21/2026.
Nvidia is hyping up its new Vera Rubin chip system this week, revealing new performance benchmarks for the GPU and CPU combo ahead of rival AMD’s annual product event in San Francisco on Thursday.
During a lengthy technical workshop last week at the company’s headquarters in Santa Clara, California, Nvidia executives boasted to a small group of journalists about the chip system’s increased power and efficiency capabilities. The biggest takeaway: Nvidia, which has long specialized in making GPUs, is increasingly trying to position itself as a supplier of CPUs that can power AI agents.
Vera Rubin is Nvidia’s successor to its hybrid superchip system Grace Blackwell and represents the linchpin of its near-term future powering the AI industry. It’s designed to offer one CPU for every two GPUs. In a single Vera Rubin NVL 72 super chip system, there are 36 Vera CPUs for every 72 Rubin GPUs. Nvidia is also selling the Vera CPU as a stand-alone product, and it has reportedly told Chinese customers these could be ready as soon as August.
Nvidia executives emphasized that its new Vera Rubin NVL72 racks—a stack of chips packed into a single liquid-cooled platform—are much more “plug-and-play” than some of its earlier products. During a brief tour of a Nvidia data center lab in Silicon Valley, Nvidia executives shared that OpenAI already has one Vera Rubin rack in use.
Nvidia claims that the Vera Rubin NVL72 system will process 10 times as many tokens per watt as the company’s Grace Blackwell super chip. The company says that its Vera CPU is faster at processing agentic AI tasks compared to rival CPUs from AMD and Intel (though the tests it ran to support those benchmarks appears to have used slightly older generations of its competitors’ CPUs). Localized memory subsystems on the new chips will also offer nearly three times as much memory bandwidth as Blackwell, which will likely be an appealing feature to many companies amid an ongoing shortage of high-bandwidth memory.
Ever since Nvidia unveiled Vera Rubin in the spring of 2025, the company has been slowly dribbling out more details about the chip system while insisting it will be released on schedule. Huang has repeatedly said Vera Rubin is ramping to “full production” and will ship in the second half of this year, with early customers including Microsoft, OpenAI, and Oracle.
5.
Alec Ibarra, Bryan Shan, Daniel Nishball, et al., “Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis,” SemiAnalysis, 07/23/2026.
Vera Rubin NVL72 is the second generation of Nvidia’s rack-scale Oberon architecture, and its gains on inference come from extreme co-design. Early results from engineering samples are encouraging. Vera Rubin NVL72 running DeepSeek R1 delivers 5.4x performance per MW and 5x performance per dollar over GB200 NVL72 today, and the gap is even wider against GB200 NVL72 during its early bringup in 2025. Vera Rubin is still in the early bringup stage now, so we expect the gap to continue widen. Rubin’s inference performance will keep improving as software matures, the same pattern we demonstrated for Blackwell in our InferenceX benchmarks, and Rubin still has a long runway ahead.
Nvidia has also recently made available their first public release of the Rubin (SM_107) software stack with CUDA13.4 and has upstreamed Rubin PRs to PyTorch, vLLM and OpenAI Triton Compiler. Blackwell was not able to reuse Hopper WGMMA kernels, Rubin is able to reuse Blackwell’s kernels, which makes the software bring up process much smoother. For speed of light (SOL) performance, engineers will still need to tune and rewrite kernels but for those that are focused on time to market, Blackwell kernels can be reused. We will also explain Rubin’s new 3 bit programmable LUT tensor core.
The early metrics gathered on VR NVL72 come from CoreWeave. We have not independently verified them. Nvidia has committed to submitting verifiable numbers to InferenceX by Q3 CY2026. Google should submit TPUv7 results in the next couple of months, and AMD has committed to MI455X UALoE72. Once those land, the ecosystem gets an objective comparison across systems.
In this article we break down Nvidia’s Rubin claims against several baselines, showing where Rubin clearly leads Blackwell and where the lead is thinner. We will also analyze Rubin’s performance per total cost of ownership using our already existing estimates for Rubin’s total cost of ownership (TCO). The TCO for Rubin and many other systems is sourced from our AI TCO model, which tracks the total cost of ownership of different AI chips, factoring in capex, opex and different other expenses. We also consider performance per watt using our All-in Utility Provisioned Power Estimates from our Datacenter Model.
Finally, we will present a component by component build up of the Bill of Materials (BoM) for the VR NVL72. This is available in our upcoming SemiAnalysis Bill of Materials (BoM) Model.
6.
Joel Khalili, “This Former Intel CEO Wants to Jumpstart Moore’s Law With Light,” WIRED, 07/21/2026.
Gelsinger announced that he had taken a position as general partner at Playground Capital, a venture capital firm specializing in deep tech, taking bets on fledgling technologies built on new science.
Gelsinger took the role with the aim of helping a new generation of semiconductor startups reawaken Moore’s law. Decades ago, Intel cofounder Gordon Moore predicted that the number of transistors on a chip would double roughly every two years, with a corresponding leap in performance. That held true for a long time, but as semiconductor firms press the limits of physics, further shrinking atomic-scale transistors has become prohibitively difficult and expensive.
Gelsinger believes the best way to break that impasse is through advances in lithography—etching chips using nanometer-scale beams of light. The leading lithography technology, developed by the Dutch company ASML, lets chip manufacturers print using 13.5-nanometer wavelength light. The ability to cram in more, even smaller features, the logic goes, would unlock even more powerful processors.
When he joined Playground, Gelsinger took a board seat at xLight, a portfolio company developing novel lithography techniques that also recently received investment from the US government.
What about memory?
High-bandwidth memory is a problematic technology. It’s the best we have right now. But by the end of the decade you’ll start to see stacked memory architectures will become much more dominant. d-Matrix, Fractile, and Cerebras are breaking the boundaries of what memory architectures will look like.
I believe that some of these hardware innovations are going to be 10 to 100X better. That means one gigawatt produces 10 gigawatts worth of tokens. I’ll take that deal any day.
7.
The Economist, “Does SpaceX’s plan for a data centre in space add up?,” The Economist, 07/22/2026.
With every passing day, it seems like a slightly less crazy idea to put data centres into space. Progress in artificial intelligence (AI) is increasing demand for these facilities, but opposition to building them on Earth is growing.
Starlink, his firm’s satellite-internet system, has already demonstrated much of the needed technology, he said. But how feasible are such claims? In the video, Mr Musk and other SpaceX executives outlined the design of AI1, the satellite that would be the building-block of Starmind, the company’s proposed orbital data centre. Unlike its communications sibling, an AI satellite does not need fancy radio antennas to communicate with the ground. What it does need is a lot of solar cells to power the AI chips at its core, and a large radiator to dissipate the resulting heat.
SpaceX’s AI1 satellite will have two wing-like solar arrays on either side of the core, each of them 30 metres by 10 metres in size, delivering a peak power of 150 kilowatts (kW). This is a bit larger than the solar arrays on the forthcoming Starlink v3 satellite, which are 19 metres long. The solar cells will deliver 250 watts per square metre, a plausible figure, and will be manufactured by SpaceX at its facility in Bastrop, Texas. SpaceX plans to put its AI1 satellites into special orbits that keep them in daylight 98% of the time.
Then there are the chips. The AI1 reference design, with a peak power of 150kW, can accommodate 72 of the most advanced AI chips made by Nvidia (equivalent to a single server rack in a terrestrial data centre). But other chips could also be used, and SpaceX plans to start manufacturing its own at a giant factory called Terafab. Finally, optical data links will connect each AI1 satellite to its neighbours in the constellation, so that hundreds or thousands of them collectively form an AI data centre. The existing Starlink system will provide connectivity to the ground.
The prospects for orbital data centres also depend on more down-to-Earth factors. Will demand for AI computing capacity continue to grow? An industry crash or a technical breakthrough could slam on the brakes. And will opposition to terrestrial data centres intensify or subside? Orbital data centres may not violate the laws of physics. But the economics and politics are less clear-cut.
–