Weekly: DeepSeek constrained by compute

14 min read.

Highlights

1. AI bottlenecks abound. The Economist offers a sort of state-of-the-AI-industry essay, listing the key bottlenecks that are hitting AI. They include: political constraints as data centres become broadly unpopular, not enough GPUs to fill expected data centres, not enough memory chips, not enough CPUs, and not enough investment to build out capacity along the supply chain.

2. DeepSeek’s compute constraints. A piece in the Economist and CFR both noting that the latest DeepSeek model - the largest development since the company crashed into the scene early last year - is not competitive with US models. CFR notes that the DeepSeek model is “is worse than leading U.S. models, more expensive than its Chinese competitors, and not able to be deployed at scale due to compute shortages.” It is reportedly modelled on Huawei’s Ascend chips, though the US government also claims that it was retrained on smuggled Nvidia Blackwell chips. Regardless, chips are evidently emerging as a key constraint on Chinese AI model development, suggesting that US AI chip export controls are achieving their intended effect.

3. Data centre backlash. A good deep, close-to-the-ground investigation by the FT on the popular and bipartisan backlash against data centre construction in the US. Public concerns include the effect on the environment, energy, water, quality of life in nearby communities, industrialisation of American farmland, and more. While 87% of existing data centres sits in urban regions, 67% of planned data centres will be in rural areas.

Thanks for reading.

---

Table of Contents

1. The Economist, “The AI rush is hitting a bottleneck,” The Economist, 04/27/2026.

2. The Economist, “Why DeepSeek’s sequel failed to impress,” The Economist, 04/28/2026.

3. Chris McGuire, Michael Horowitz, and Jessica Brandt, “DeepSeek V4 Signals a New Phase in the U.S.-China AI Rivalry,” CFR, 04/29/2026.

4. Susannah Savage, Rafe Rosner-Uddin, Eva Xiao, and Zehra Munir, “The great American data centre divide,” FT, 04/28/2026.

5. Asa Fitch, “Musk’s Chip-Making Vision With Intel Is a Distant Prospect,” WSJ, 04/24/2026.

6. Daniel Nishball, Dylan Patel, Cheang Kang Wen, et al., “AI Value Capture - The Shift To Model Labs,” SemiAnalysis, 05/01/2026.

7. Rishi Iyengar and Christina Lu, “The Hormuz Hit to Helium,” Foreign Policy, 04/27/2026.

---

1.

The Economist, “The AI rush is hitting a bottleneck,” The Economist, 04/27/2026.

Start with the politics. In April legislators in Maine voted in favour of a bill to ban the construction of data centres above 20 megawatts until November next year. Although it was subsequently vetoed by the governor, lawmakers in more than ten other American states are weighing similar measures. According to one count, $156bn-worth of data-centre projects were blocked or delayed last year in America by local opposition and litigation. Other countries, from Ireland to Brazil, are experiencing a growing backlash. Concern over the impact of power-hungry data centres on electricity bills in particular has become widespread—and may intensify further as the war in the Gulf raises energy prices.

Ivan Chiam of SemiAnalysis, a research firm, points out that there are not enough chips to fill the data centres now being built. Consider the graphics-processing units (GPUs) designed by Nvidia, which provide more than two-thirds of the world’s AI computing power. The price to rent one of its H100 GPUs, launched in 2022, has soared by around 30% since November, as customers unable to get their hands on newer models have resorted to older generations. Competing AI processors are also getting more difficult to obtain. In April Andy Jassy, Amazon’s boss, said that his company had nearly sold out access to its Trainium2 AI chips. A significant chunk of the capacity of Trainium4, due next year, “has already been reserved”.

The squeeze also extends to memory chips, in particular the kind of high-bandwidth memory (HBM) that AI models rely on. All three big producers—SK Hynix, Samsung and Micron—say that most of their supply for 2026 is sold out. Some hope of relief came in March when Google unveiled TurboQuant, an algorithm meant to reduce the amount of memory AI needs, causing the share prices of the memory-makers to briefly swoon. Even so, demand for HBM is expected to outstrip supply for at least the next three years.

The shortages are now spreading to central-processing units (CPUs). “Agentic” AI tools, which plan, reason and carry out tasks, rely more heavily on these types of chips to co-ordinate their work. Morgan Stanley, an investment bank, estimates that agentic systems require one CPU for every GPU, compared with a ratio of one to 12 for chatbot-style systems. Indeed, demand for CPUs has been so robust that it has breathed new life into Intel, which not long ago seemed to be heading for collapse. The market capitalisation of the American chipmaker, one of the leading producers of CPUs, has more than doubled over the past six months (see chart 1).

The crux of the problem is that companies along the AI supply chain are investing far less than the hyperscalers. We examined the planned capital spending this year of the 50 or so largest manufacturers of chips, chipmaking tools, servers, networking gear and cooling equipment, and how it has changed since 2024. The five hyperscalers have tripled their combined capital spending, to more than $750bn, but the hardware suppliers have increased theirs by only half, and will invest less than a third as much as the cloud giants this year (see chart 2).

2.

The Economist, “Why DeepSeek’s sequel failed to impress,” The Economist, 04/28/2026.

A little more than a year ago a small Chinese artificial-intelligence startup shocked the world. When DeepSeek unveiled a pair of models that performed nearly as well as the best Western ones, but were built for a fraction of the cost, a panic swiftly followed. The share prices of Nvidia and other providers of AI infrastructure briefly tumbled as investors fretted (wrongly, it turned out) that demand for their wares would slow in the face of such a leap in the efficiency of model-making. Yet the release on April 24th of the lab’s new model, called v4, has been greeted with a shrug. Why?

It seems that, unlike DeepSeek’s previous blockbuster, v4 was not particularly cheap to build. In 2025 the lab eagerly pointed out that the cost of training its AI was about $6m, far below the going rate in the West. The lab’s technical white paper on v4, however, omits any estimate of this measure. The fact that 16 months elapsed between v4 and its predecessor also hints that oodles of processing power were used to train it.

The release comes at a time when China’s AI scene is increasingly crowded. DeepSeek has faced growing competition both from other independent labs, such as Moonshot and Z.ai, and the country’s internet giants. The Qwen family of models produced by Alibaba, an e-commerce colossus, has sat comfortably atop China’s leader-board for most of the past year. ByteDance, creator of TikTok, is also the maker of Doubao, China’s most popular chatbot. Dola, as it is called outside China, is hugely popular in Mexico, the Philippines and Britain, where it ranks above Google’s Gemini in Apple’s app store.

At the same time, DeepSeek has had to contend with greater state meddling at home. China’s government has been promoting chips made by Huawei, the national semiconductor champion. DeepSeek reportedly tried to train its new model on them, but eventually fell back on Nvidia’s chips instead, adding cost and time. The government seems unlikely to give local AI companies a freer hand soon: on April 27th it said that it would block the acquisition of Manus, another of the country’s AI darlings, by Meta, an American social-media giant. The startup’s co-founders have been barred from leaving China since March.

3.

Chris McGuire, Michael Horowitz, and Jessica Brandt, “DeepSeek V4 Signals a New Phase in the U.S.-China AI Rivalry,” CFR, 04/29/2026.

DeepSeek has released its long-awaited new model, DeepSeek V4—its first new-architecture model since the release of R1 in January 2025. Whereas R1 stoked alarm that China was catching up to the United States in AI, and even briefly tanked U.S. stock markets, V4 has produced a more muted reaction. This new DeepSeek model is not competitive with frontier U.S. models. And while it is likely the best available open-source option, it does not provide evidence that Chinese AI firms are shrinking the gap with the United States.

DeepSeek’s own technical paper concedes that V4’s reasoning and agentic capabilities are comparable to GPT-5.2, Gemini 3.0 Pro, and Claude Opus 4.5—models released roughly half a year ago. DeepSeek even explicitly acknowledges that V4 “trails state-of-the-art frontier models by approximately 3 to 6 months.” That gap is broadly consistent with estimates that the United States has roughly a seven-month lead over China. The actual gap may also be widening, as U.S. AI firms use AI to accelerate next-generation model development. The newest U.S. models announced in April—Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5—both show significant performance gains over their predecessors.

The most novel feature of V4 is that it is optimized for inference on Huawei’s Ascend chips rather than Nvidia’s, reportedly at Beijing’s direction. U.S. government officials have asserted that V4 was still trained on smuggled Nvidia Blackwell chips, which are banned in China. It is notable that, unlike the V3 paper, the V4 report is silent on what chips it was trained on.

And DeepSeek itself admits that it currently cannot serve its v4 pro model to most customers because it lacks the chips to do so.

The result is a model that is worse than leading U.S. models, more expensive than its Chinese competitors, and not able to be deployed at scale due to compute shortages.

4.

Susannah Savage, Rafe Rosner-Uddin, Eva Xiao, and Zehra Munir, “The great American data centre divide,” FT, 04/28/2026.

Data centres, once clustered around cities and towns, are moving into farm country in search of cheap land and tax incentives. According to Pew Research Center, 67 per cent of planned data centres are in rural areas, while 87 per cent of existing data centres are in urban ones.

As the industry has expanded, public opinion has hardened against it. Pew research found that Americans are far more likely to view data centres as harmful than beneficial in terms of environmental impact, domestic energy costs and quality of life in nearby communities.

The issue is awkward for President Donald Trump and his party. Republican strategists are increasingly wary that the administration’s support for AI could trigger a backlash among key voter blocs, including farming communities, ahead of November’s midterm elections. Around 78 per cent of US counties dependent on agriculture voted for him in 2024, according to analysis of election data by Investigate Midwest.

In rural areas from Illinois to West Virginia, new data centre proposals have led to packed public meetings and organised opposition as residents push back. In Indiana, shots were fired at a local lawmaker’s home and a note left on his doorstep reading “no data centers”. Democratic politicians have called for tighter regulation and Republicans in several states have campaigned against new developments, reflecting the backlash.

But the picture is mixed for farmers. While some worry about the industrialisation of once-agrarian communities, others welcome the opportunity to cash in on soaring land prices or to generate additional income.

The debate has reverberations far beyond America’s farm belt, pitting two visions of the country’s economic development against each other. In the view of the White House and much of Big Tech, the data centres will help the US maintain its lead in AI. Expectations about the growth the infrastructure will deliver underpin everything from sky-high share valuations to state and federal tax breaks granted in the hope of new jobs and investment.

But rural communities, along with many Americans, worry about the immediate impact of data centres on water and power costs and the broader disruption they represent for people’s way of life.

5.

Daniel Nishball, Dylan Patel, Cheang Kang Wen, et al., “AI Value Capture - The Shift To Model Labs,” SemiAnalysis, 05/01/2026.

A day in AI now feels like a year in any other industry. Model releases, software breakthroughs, and hardware improvements are compressing multi-year cycles for any other industry into weeks. Over just the past few months, agentic AI has crossed a real inflection point, driving a step-change in the value of tokens while software and hardware improvements have sharply reduced the cost of generating them.

New chips such as Blackwells can generate 30x more tokens per second while running frontier workloads today vs Hoppers a year ago, and ASICs such as TPUv7 and Trainium 3 show similar improvements. Inference providers such as Fireworks, Baseten, Fal, margins are widening while their revenue trends are in hyper growth.

Even parts of the hardware stacks have repriced, with memory prices having gone up 6x in the past year. Neocloud GPU rental pricing is surging as well, up with 1-year H100 rental contract prices up 40% from the bottom in October 2025.

There are two firms in the industry with incredible pricing power that haven’t moved much though. TSMC and Nvidia have not reacted to the recent boom in value generation of AI models.

In this article we will explore where value from AI is accruing - from end users to inference providers, Neoclouds as well as hardware providers. We will unveil how TSMC and Nvidia are now venting vast value into every vertical of the ecosystem.

6.

Asa Fitch, “Musk’s Chip-Making Vision With Intel Is a Distant Prospect,” WSJ, 04/24/2026.

Some of Elon Musk’s ventures—Tesla, SpaceX—have been spectacularly successful. Others—The Boring Company, Neuralink—haven’t come close to matching his grand visions.

Musk’s latest moonshot is a sprawling, vertically integrated chip-making operation called Terafab. Don’t expect it to emerge from the latter category anytime soon. That in turn means the benefits for most companies involved, including chip giant Intel, also lie largely in the far-off future, if they materialize at all.

Musk’s targets for the Terafab have been characteristically fabulous: Mass production is supposed to start next year, with an initial target of 100,000 silicon wafers a month. If all goes to plan, it will grow to a million wafers a month.

To put those figures in context, the largest factories of the world’s largest chip maker, by market value Taiwan Semiconductor Manufacturing Co., put about 100,000 wafers a month into production. A million a month would be around 70% of TSMC’s total monthly output, per analysts at New Street Research.

What is certain is that this will take much more time than Musk imagines, even if he isn’t overestimating his ultimate needs. Chip factories require specialized construction capabilities and materials—seismic-resistant concrete, for example, that absorbs even the most minute vibrations in the Earth’s crust. Building chip factories involves a lot of location-specific engineering that can’t happen overnight.

And Musk doesn’t want to build a standard factory. He is pushing a novel method of chip-making that brings many parts of the supply chain under one roof. On a call with analysts Wednesday, he gave an indication of just how gradual the rollout actually will be. Tesla would first build a chip-research facility with capacity to produce “a few thousand wafers” monthly, spending about $3 billion, he said—chump change in the chip-manufacturing world. The intent was “to try out ideas,” he said.

7.

Rishi Iyengar and Christina Lu, “The Hormuz Hit to Helium,” Foreign Policy, 04/27/2026.

The United States and Iran may no longer be in active conflict, having established and then extended a cease-fire in their weekslong war, but the list of economic disruptions and industrial shortages continues to grow—oil, natural gas, jet fuel, tungsten, sulfur, fertilizer.

You may know helium as the gas that keeps party balloons afloat and pumps voices to new heights, but the gas has farther-reaching impacts than that. Colorless and lighter than air, helium is also a vital ingredient in many of the world’s most powerful technologies, underpinning semiconductors, medical equipment, and more.

If “you think about semiconductors, fiber optics, or anything using superconducting magnets, the consequences of not having helium are gigantic in an economic sense,” said Nicholas Snyder, the CEO of North American Helium, which currently produces more than 7 percent of the helium supply in North America. “If you think about medical applications like MRI, there [are] huge consequences there as well.”

Weeks of war in the Middle East and attacks on energy infrastructure have dealt a painful shock to the global trade of helium, which is primarily extracted as a byproduct of natural gas production. As the world’s second-biggest producer of liquefied natural gas (LNG), Qatar in particular is a major helium hub, accounting for about one-third of global helium supply before the war erupted.

That changed after Iran attacked Qatar’s Ras Laffan facility—the biggest LNG plant in the world—prompting the state-owned QatarEnergy to halt production, declare force majeure, and slash its annual helium exports by 14 percent.

Exports are also having trouble leaving the region: Most of Qatar’s helium exports normally transit the Strait of Hormuz, the vital maritime chokepoint that has been effectively strangled by the war.

The turmoil has already spiked helium prices, with spot prices reportedly doubling last month. But since the commodity is mostly traded through long-term contracts—and the global transportation of helium containers entails a longer lag time—more pain could be on the horizon.

This is a “belated wake-up call in the helium world,” said Snyder, particularly for places such as South Korea and Taiwan, he said, which were getting the majority of their helium from one source in the Middle East.

–

Cover Photo by Solen Feyissa on Unsplash