Weekly: Visualising the AI buildout + Personal News
21 min read.
(Repost from today’s Daily) Personal news: I recently left my position as a policy consultant for the World Bank. I’m grateful to have worked on semiconductor and technology projects for the government of Vietnam this past year. Next month I will be starting a JD at Stanford Law School and a PhD in the Department of History at Stanford University. I’m fortunate to be funded by the Knight-Hennessy Scholars program.
I’ve been writing Chip Briefing every weekday for the past two years now - more than 400 editions. I’ve learnt a lot and it has been good to connect with readers. I would like to continue the newsletter in its current form and keep it free too. But writing this newsletter, on top of other commitments, takes a lot of work. I’m told law school, especially the first year, is rather busy. I am talking to friends and mentors, and thinking hard about what Chip Briefing could look like going forward. Thoughts are welcome through the poll below, reply email, or at chipbriefing@substack.com. In the meantime, if you find the newsletter helpful and would like it to continue, please share it with friends and colleagues.
Highlights
Visualising the AI buildout. The NYT has an interactive, graphic story on the absolutely tremendous scale of the AI buildout. Finding good information is a bit tricky but they report that the number of AI is expected to “double roughly every nine months,” meaning there would be some 200 million AI chips in the world by the end of 2028, 10x what we have now.
China’s chip self-sufficiency. The WSJ has an in-depth piece on China’s AI chip buildout. It is by far the main constraint to China’s AI ambitions. Ding Xuexiang, the very powerful Vice Premier of China, is in charge of China’s chip self-sufficiency drive. Still, analysts say that China remains behind the US in both chip quality and quantity, and will likely stay behind through 2030.
Complementary pieces in the FT. One by Chip War author Chris Miller who argued against allowing the sale of CXMT chips to US firms. Another, a profile of ByteDance, which is “the best-positioned tech giant in China in terms of its raw computing capacity.”
Why chip stocks are down. The Economist writes on why South Korea stocks are dramatically down. South Korea, whose population has a particular retail fervour, poured absurd volumes of money into single-stock leveraged ETFs. Even small selloffs got amplified and now the government is tightening regulations around them.
FT Alphaville writes that chip stocks are in decline as part of their historical cyclicality. For consumer electronics customers of memory chips, the chips are becoming too expensive. The raised prices are denting demand for phones and so, denting demand for chips.
Thanks for reading.
Table of Contents
Adam Satariano, Paul Mozur, Jacqueline Gu and Cade Metz, “The Impending, Inescapable Deluge of A.I.,” NYT, 07/29/2026.
Josh Chin and Raffaele Huang, “Inside China’s All-Out Push to Catch Up With American AI Chips,” WSJ, 07/23/2026.
Chris Miller, “China is not the solution to the US chipflation problem,” FT, 07/31/2026.
Zijing Wu and Eleanor Olcott, “ByteDance’s big bet on AI,” FT, 07/30/2026.
Lex, “Nvidia’s chips may be novel, but its ‘circular financing’ isn’t,” FT Lex, 07/30/2026.
The Economist, “South Korea’s stock-market boom is collapsing spectacularly,” The Economist, 07/27/2026.
FT Alphaville, “Maybe the chips are down because memory’s too expensive,” FT Alphaville, 07/28/2026.
Lex, “CXMT’s roaring IPO isn’t the bubble signal it might appear,” FT Lex, 07/28/2026.
Mari Kiyohara, Mayumi Negishi, and Takashi Mochizuki, “Race for Chip Riches Heats Up as Kioxia Faces Emboldened Rivals,” Bloomberg, 07/29/2026.
Bryan Shan, Daniel Nishball, Myron Xie, et al., “Can AMD break the CUDA Moat? AMD Advancing AI 2026,” SemiAnalysis, 07/25/2026.
1.
Adam Satariano, Paul Mozur, Jacqueline Gu and Cade Metz, “The Impending, Inescapable Deluge of A.I.,” NYT, 07/29/2026.
The milestones for artificial intelligence keep getting grander. Behind each leap in A.I. are corresponding jumps in computing power. Today, there are about 20 million A.I. chips crammed into the data centers that underpin the technology’s growing abilities and usage worldwide, according to the research firm Epoch AI. That figure is expected to double roughly every nine months, putting the world on pace to have about 200 million of the chips by the end of 2028 — 10 times current levels.
In size and ambition, this moment compares to the building of the railroads in the 1800s, President Franklin D. Roosevelt’s New Deal in the 1930s, and the Manhattan Project to create an atomic weapon in the 1940s, technologists said.
But the build-out has also stoked a backlash, spurring protests in many communities over how data centers could harm the environment, raise electricity prices and drain water. In the United States, data centers are shaping up as a major issue for November’s midterm elections, with a growing national movement pushing back against the tech industry and its billionaires.
Lack of transparency in the A.I. industry makes measuring global computing capacity difficult, including the volume of chip supplies, total number of data centers and overall electricity consumption. The New York Times relied on estimates from groups including Epoch AI, Cleanview and SemiAnalysis that study the industry and publish widely cited forecasts.
For now, there are no signs that the spending on A.I. will slow. By 2029, A.I. infrastructure investment is forecast to top $1 trillion globally, up from $318 billion last year, according to IDC, the market research firm. That would be on par with the economic output of Switzerland, the world’s 20th-biggest economy.
China, the next closest rival in A.I., is working to close the gap. Chinese companies had roughly 1.16 million H100-equivalent chips at the end of 2025, up from roughly 244,000 at the beginning of 2024, though those figures exclude smuggled chips and other offshore computing resources used by Chinese firms, Epoch AI estimated.
China’s top tech companies are building new A.I. chips and data centers. This year, Huawei, ByteDance and Alibaba are expected to spend $111 billion on data centers and other A.I. investments, according to Bernstein Research.
2.
Josh Chin and Raffaele Huang, “Inside China’s All-Out Push to Catch Up With American AI Chips,” WSJ, 07/23/2026.
In May, Huawei went public with more details of its plans. The company already led the way in reducing China’s dependence on foreign AI chips from 90% in 2021 to less than 60% by 2025, and the newer designs could help lower that number to 25% over the next half-decade, according to data from Morgan Stanley. What’s more, Huawei said it had figured out workarounds for making near state-of-the-art silicon without leading-edge machinery.
The outcome of China’s effort to catch up has ramifications for the global economy. If the U.S. controls the top technology, the rest of the world will depend on Washington’s favor to make progress in fields such as medicine, robotics and autonomous driving—not to mention weaponry and military strategy. AI advances depend on hardware, in particular the chips that train AI models and help them tackle everyday tasks.
A record of overtaking the West in electric vehicles and batteries gives many in China confidence the country can eventually break the U.S. stranglehold on chip technology. But chips are different from other hardware. The technology is extraordinarily complex, and it is moving at a lightning pace.
Ding’s committee bets that by hitching China’s engineering prowess to the engine of state capitalism, it may eventually get there. The push has made progress that few thought possible back in 2022, when Huawei was starved of U.S. technology and Chinese AI companies depended almost entirely on Nvidia.
Executives at Chinese tech companies described a change in mindset over the past year. They had previously dismissed local chips as impossibly far behind. Now they think they can do business with Huawei and other local chip makers. Tech company Meituan recently released an AI model that it said was comparable to Google’s flagship model and was trained solely with Chinese chips.
In memory chips, another key AI need, Chinese companies that barely had sales a few years ago can now catch a glimpse of the American and South Korean companies at the front of the pack.
Chinese chip makers aim to boost the production of advanced wafers used to make chips to more than 500,000 a month by 2030, up from roughly 30,000 last year, according to people familiar with the target. Huawei expects to ship around 1.5 million AI chips this year, roughly doubling its 2025 volume, some of them said.
Still, the technology gap remains considerable. Nvidia’s top AI chip boasts around four times the computing power of Huawei’s best offering. China’s AI computing power stood at around 14% of the U.S. in 2025, research house Bernstein estimated. Despite China’s aggressive build-out, Bernstein expected China would remain significantly behind the U.S. through 2030.
Under Ding’s watchful eye, Chinese companies are using hundreds of thousands of Nvidia chips to stay within striking distance of their American rivals for now.
3.
Chris Miller, “China is not the solution to the US chipflation problem,” FT, 07/31/2026.
The Trump administration, nervous about inflation amid the energy shock of the Iran war, is looking to bring prices down. One option is to buy more chips from China. For many years China was highly dependent on western companies for memory, but this has now changed. This week saw the initial public offering of Hefei-based CXMT, which manufactures DRam, one of the two main types of memory chips. It briefly became China’s most valuable company, a staggering rise for a company recently seen as second tier.
Even so, the US government has discouraged western companies from using CXMT chips. Until recently, many steered clear. But with memory in severely short supply, they are now reassessing their options. Apple, for example, reportedly asked for the White House’s blessing to use CXMT chips. It is a short-sighted solution.
For the moment, most of CXMT’s production is commodity DRam, not the ultra-complex high-bandwidth memory used to train AI systems. If the company breaks into the leading edge it could reshape the memory industry.
Some buyers of memory chips see CXMT as the answer to chipflation. This seems unwise. But global peers worry that the company, which is roughly half owned by Chinese-government-linked entities, could build capacity without concern for profitability, making buyers dependent on Chinese supply. At stake is not only short-term prices for memory but an industry that has never been more strategic.
4.
Zijing Wu and Eleanor Olcott, “ByteDance’s big bet on AI,” FT, 07/30/2026.
Over the past three years, ByteDance has accordingly invested in AI more aggressively than any of the other Chinese tech giants, building out its network of data centres and hiring researchers. It has doubled down on its cloud unit, Volcano Engine, which sells AI solutions to enterprises, and buttressed its video generation capabilities with its model Seedance. It also has ambitions to develop a custom AI chip.
There is a broader, deepening investor optimism about China’s advances in AI, particularly with the latest models from Moonshot AI and the blockbuster initial public offering of memory-chip maker CXMT. The value of ByteDance’s shares on the secondary market has risen to up to $600bn, a 50 per cent jump from last year.
But investors and analysts say ByteDance’s metamorphosis represents a big gamble at a moment when some are questioning the vast sums being spent on AI. Semiconductor stocks have plunged in recent weeks over concerns about oversupply and the sustainability of the AI boom.
Past deviations from this core business have not always worked out, including previous forays into gaming and education technology. Nonetheless, ByteDance is now ploughing proceeds into AI infrastructure, hardware and software as it raises money to expand into new areas such as chips.
ByteDance formulated a plan to buy more Nvidia cards for AI training, while Zhang personally oversaw a hiring spree for top research talent.
Founded in 2020, ByteDance’s cloud unit Volcano Engine was a new entrant in an already saturated market where Tencent, Baidu, Huawei and Alibaba all offer business services, in addition to state-owned cloud companies such as China Mobile. ByteDance has gained market share by offering cut-throat pricing on its fleet of products and models.
Another bottleneck for all Chinese AI companies is the lack of advanced computing power in the country.
ByteDance is the best-positioned tech giant in China in terms of its raw computing capacity, having been an early and aggressive purchaser of Nvidia chips as well as domestic inference chips from Cambricon, say industry insiders.
5.
Lex, “Nvidia’s chips may be novel, but its ‘circular financing’ isn’t,” FT Lex, 07/30/2026.
The flurry of criss-cross transactions and partnerships among AI chipmaker Nvidia and its customers has created much consternation about circular financing. The term itself is ominous, and the companies themselves reject it. After all, circularity suggests something that ultimately goes nowhere.
But for all the fancy ways in which Nvidia is extending its support to the likes of OpenAI and data centre builders that use its chips, what’s really happening is old-fashioned vendor financing. Like telecoms equipment manufacturers, planemakers and department stores of old, the chipmaker is basically writing cheques to enable its customers to buy more of its products than they could otherwise afford.
These transactions are not actually circular — which would suggest no value is actually created — but back to back. Equity investments aren’t necessarily funnelled directly back into buying chips. And even where they do result in a chip sale, nothing is given free. Nvidia gets the revenue, and still has a stake in a company.
Another spin on the plump-up-your-customer theme is the credit guarantee. Nvidia already backstops loans for some customers, totalling $3.5bn at the last count, but may soon offer support to OpenAI amounting to $250bn, according to the Wall Street Journal. The company would have to disclose the gross amount, but if structured as a credit derivative, it may not make much of an impact on its balance sheet.
But vendor financing is also a double blow when things go wrong: one loses one’s customer and one’s financial exposure. Credit guarantees ratchet up risk. Whereas a badly judged equity stake might later prove to be a waste of money, the requirement to take on a customer’s debts if things go awry can turn a valuation problem into a solvency problem.
6.
The Economist, “South Korea’s stock-market boom is collapsing spectacularly,” The Economist, 07/27/2026.
Now a nation newly enamoured of stock markets is discovering that shares can go down as well as up. Since peaking in June the stock prices of Samsung and SK Hynix have crashed by 41% and 52% respectively, wiping $1.2trn from South Korea’s highly concentrated stock market. Those of SK Hynix fell by more than a fifth this week as investors digested a mere six-fold increase in its quarterly operating profit, compared with the year before. Samsung’s shares have not been this volatile since the dotcom era.
South Korea is a great experiment in mixing the extremes of state capitalism and market speculation. Last year, after a failed coup, the country’s president attempted to cheer his country by talking up the stock market (sound familiar?). The KOSpI, its main index, speedily hit a level the government had set as a target.
Ordinary South Koreans withdrew capital from insurers and banks to invest in stocks. Workers threatened industrial action to secure Wall Street-sized bonuses. Those at SK Hynix and Samsung’s chip division settled for an extraordinary 10% of operating profits. At SK Hynix that could amount to $500,000 a person this year, and $800,000 next year. (Goldman Sachs paid a meagre average of $400,000 in total compensation to its employees last year.)
Now a government that just weeks ago was embracing markets is being forced to apologise for them. It was only in April that regulators approved leveraged exchange-traded funds (ETFs) tied to individual South Korean stocks. These funds use derivatives to give investors a multiple of the daily return of a share. For example, if SK Hynix is up by 1%, an investor might receive 2%; or if it is down by 10%, an investor might jump out of a window.
When the selling began in June broker accounts faced margin calls and were shut down by the thousand, causing regulators to suspend new ETF launches. At an emergency meeting on July 29th politicians promised to severely limit the use of these funds. One compared the market to a casino. Shares in the country’s department stores are crashing as fast as those of its chipmakers, in anticipation of the handbags that will now remain on the shelves.
The trillion-dollar question for investors in South Korea is when the price of memory will peak. Bulls argue that data-centre construction and robotics mean demand for memory will outstrip supply into the 2030s. But the increasingly common view is that the same thing will happen to memory chips as has happened to every commodity in every commodity cycle in history—that new supply will, within a year or two, cause prices and profits to fall. Despite their vertiginous rise, Samsung and SK Hynix have on average traded only at single-digit multiples of forward earnings during the past year, suggesting that investors think their profits are at or near their highest point.
7.
FT Alphaville, “Maybe the chips are down because memory’s too expensive,” FT Alphaville, 07/28/2026.
Whoever could’ve predicted that the year’s boom for the most notoriously cyclical sector might reverse on concerns about cyclicality?
July’s chip dip looks most like an unwind of a crowded trade, though it’s a new reason to talk about unsustainable big-tech capex, debt, overbuild, token deflation, Chinese competition, etc. A few reports pin the chip pullback on a single-sourced story from Reuters about a state-owned Chinese company beginning mass production of immersion deep-ultraviolet lithography machines, while several more cite the IPO of CXMT in Shanghai. Sure, fine.
The more obvious concern for now is demand destruction.
Industry optimism has weakened on signs that only the hyperscalers and AI infrastructure providers are willing to pay top dollar for memory. If that’s correct, it would suggest chip prices are plateauing sooner than expected and revenue forecasts have to come down by a lot.
Consumer electronics makers have struggled to pass through last quarter’s sequential price increases for memory of 80 to 100 per cent, say Jefferies analyst Edison Lee and team. Smartphone and gadget makers are pushing back amid very weak demand, so market expectations for prices of DRAM and NAND to rise a further 25 to 30 per cent this quarter look overly optimistic by about 10 percentage points, they say.
What unravels as chip prices plateau is the circularity of the trade. Hyperscalers have raised capex guidance to spend more on chips, which in turn has squeezed chip prices higher. All that’s needed to wreck 2027 sales expectations is for this cycle to turn from virtuous to vicious. Plus ça change . . .
8.
Lex, “CXMT’s roaring IPO isn’t the bubble signal it might appear,” FT Lex, 07/28/2026.
CXMT’s initial public offering is a great illustration of what happens when tokenmaxxing meets Chinamaxxing. The chipmaker’s shares rose 466 per cent on its first day of trading in Shanghai, briefly making it China’s most valuable company.
On one hand, it couldn’t really have been otherwise. CXMT sold a tenth of its post-IPO share capital; sure, that’s twice as much as Elon Musk’s SpaceX sold in its June debut, but the $8.5bn CXMT raised is only a fraction in monetary terms. And since Chinese regulators generally don’t like expensive listings, underpricing is commonplace. Semight Instruments, a chip testing company, rose nearly 900 per cent on its debut earlier this year; the shares have risen further since.
It is worth asking whether CXMT would have done so well without those quirks. Certainly, memory chips are in high demand, and CXMT has sprung from virtually nothing to an 8 per cent market share, according to Counterpoint. But chips are a boom-bust business. Overall, consultancy Gartner sees capacity increasing by a quarter in the two years to mid-2029.
At the same time, there are nascent signs of peaking demand. Companies that rushed gleefully into using AI are starting to wince at the cost of computing tokens — units of text or text used in processing by large language models. IBM is one that has sounded the death knell for tokenmaxxing, warning that companies will instead seek to plan their AI strategy with an eye on value, rather than just trying to jack up usage.
Cycles cannot be held at bay forever, but it is true that the AI boom driving chip demand may follow a different path. New uses and configurations of hardware are emerging, which keep supply chains busy and create demand in new places. Chipmakers are turning to specialisations, too. In the US, Micron has turned its focus to servicing hyperscalers with high-bandwidth memory. Qualcomm has developed an alternative architecture that packs in more memory for less cost.
9.
Mari Kiyohara, Mayumi Negishi, and Takashi Mochizuki, “Race for Chip Riches Heats Up as Kioxia Faces Emboldened Rivals,” Bloomberg, 07/29/2026.
These days, a global spending spree in AI hardware has reversed Kioxia’s fortunes. Demand is booming for flash memory chips, which provide storage essential for OpenAI’s ChatGPT and other AI models. Kioxia’s shares became 2025’s top performer on the global MSCI World Index, with their value briefly surpassing Toyota Motor Corp.’s earlier this year. On Friday, Kioxia is expected to report a 30-fold jump in quarterly operating income that exceeds what it earned in the entire fiscal year ended in March.
Yet this dramatic turnaround — which netted Bain around $15 billion in Japan’s largest private equity exit, according to people familiar with the matter — is already being tested. Kioxia shares fell 18% on Tuesday in the latest wave of profit-taking across memory-related stocks, driven by growing concerns over whether big tech’s spending rush will continue. The Japanese company must now prove it’s more than an accidental beneficiary of an AI-driven supply squeeze.
Kioxia faces mounting pressure from customers and shareholders alike to expand capacity and defend its technological lead against SK Hynix Inc. and Samsung Electronics Co. in high-end flash memory for data centers. In the consumer segment, smaller low-cost players like China’s Yangtze Memory Technologies Co. are aggressively ratcheting up production.
The question is whether Kioxia can sustain its momentum in an industry where pricing power historically stems from massive capital commitments as well as innovation. Having endured years of financial strain and austerity, the company risks playing it too safe should rivals pour aggressive capital into the AI boom. Its predicament reflects a broader reality across the tech supply chain: The cost of keeping pace in the AI arms race keeps growing, even as doubts persist about long-term demand. A Kioxia spokesperson, asked for comment, said the company continues to make “appropriate” capital investment in line with customer needs and demand trends.
10.
Bryan Shan, Daniel Nishball, Myron Xie, et al., “Can AMD break the CUDA Moat? AMD Advancing AI 2026,” SemiAnalysis, 07/25/2026.
Based on our experience on AMD software stack this year, we update our view again from non-zero percentage chance to now a great chance of success as long as AMD solves the two major risks we outline below. It is important to highlight that just because AMD gains market share, that doesn’t mean that Nvidia will do poorly. The pie is growing rapidly for everyone, and Nvidia will continue to massively grow revenue. AMD poses potential competition to Nvidia on the software front, and Jensen will need to cut bureaucracy and flatten the layers of different required internal stakeholder approvals required for even the simplest of tasks if he wants Nvidia to move faster and defend their lead.
Anthropic has publicly announced that they will deploy 2GW of AMD’s chips, Lisa Su and team have leaned into an agentic oriented engineering culture. Anthropic Head of Compute Tom Brown explained how he used Claude over the weekend with “/goal” to bring-up internal Claude inference stack on AMD hardware as a case in point. We believe that since AMD’s compiler and most of their kernels are open sourced, that they are better positioned for the agentic age besides one major risk that we will discuss later. Three months ago, our Accelerator model noted that Anthropic will be an AMD customer. We also noted this on our socials a week ago.
In 2023, Microsoft dropped AMD after the MI300X due to unreliable Samsung 2023 HBM memory and poor software quality, subsequently skipping both the MI325X and the MI355X. Microsoft has since made an about face and has announced that it will deploy MI455X Helios. We believe that OpenAI will be the main end customer for Azure’s MI455X racks. In a strategy somewhat similar to the Nvidia-Groq deal, AMD is announcing a deal with Cerebras to do PD disagg for ultra-fast interactivity inferencing.
There are two major risks that AMD needs to guard against:
Supply chain checks and engineering first principles analysis and understanding show that AMD’s first AI rack scale system, Helios, is going currently going through a slow rack production ramp given that it is not using a cableless tray design, which the Rubin Oberon rack is now adopting. Furthermore, since AMD has a weak SerDes design, up to 85% of its backplane needs to be retimed, requiring over over 550 Broadcom ethernet retimers per rack. Furthermore, it is running into backplane reliability challenges during rack production ramp hell.
The chief complaint from most AMD engineers internally is that there is a persistent lack of stable GPU clusters for internal software development teams and a lack of stable GPU clusters for automated testing CI. This is blocking AMD’s rate of progress and it is holding AMD back from harnessing the potential upside of AI coding Agents because each AI Agent requires GPUs as well and also requires a testing tool use loop. We will more explain below on our recommendation section.
If AMD can overcome these challenges, we strongly believe that AMD will be well positioned to do well and take market share. This will be helped along as well by the recent stock option based structure whereby AMD gives Meta and OpenAI close to a 105% equity rebate discount using some clever financial engineering. The full rebate triggers a AMD stock reaches the final level of $600 and once OpenAI/Meta buys enough compute. Helios performance per TCO is so great that AMD that cost per million tokens is practically negative cost when combined with this structure! AMD is practically giving away Helios racks and an 5% extra on top of that to an SF-based nonprofit called OpenAI.
–