The AI Pipeline, Part 5: Nobody Owns the Stack, Everyone Owns a Switch

The short answer: the AI pipeline is a stack of about seven layers, minerals, design software, chip architectures, lithography, fabrication, memory, and cloud, and every layer is a monopoly or a triopoly. Nobody controls the whole stack, everybody controls a kill switch, and every switch has been pulled at least once since 2022.

This is part 5, the last part of the AI Pipeline series. Part 1 covered the minerals and the design software. Part 2 covered the lithography machine, the foundry, and the memory. Part 3 covered the buyers and the export control era. Part 4 mapped the geography of compute and walked through India's position in full. This part does the synthesis: strip away the details and say what the pipeline actually is, what the evidence says about whether it is producing anything real, who holds the kill switches, and what is actually available to the countries that own none of it.

What the pipeline is#

AI runs on a stack of about seven layers, and each layer is a monopoly or a triopoly. Sand and minerals, with China holding 70% of rare-earth mining, 90% of refining, and 98% of gallium. Chip design software, three companies holding about 70%. Chip architectures, the x86 duopoly and ARM. The machines that print chips, one Dutch company with one German optics supplier whose prolonged absence would halt production. Fabrication, one Taiwanese company at 70-73% of foundries and over 90% of advanced AI chips, plus one Japanese-adjacent packaging and materials cluster where a food company's film and a glass-cloth maker each hold around 90-95% of their niches. Memory, three companies with South Korea at 75-85% of next-generation HBM. Cloud and compute, three American companies at roughly 70%, five at 71% of AI compute. Models, a handful of labs, mostly American, one serious Chinese counterpart, with 59 of the notable 2026 models American and 35 Chinese.

Nobody controls the whole stack. Everybody controls a switch. The United States demonstrated its switches on Russia in 2022 and on China continuously since, from the A100 license requirement to the H20 write-off to the TSMC waiver pull. China demonstrated its on minerals from 2023 onward, and is building its way out of the one switch it lacks (lithography) with DUV production running now and EUV prototypes targeted around 2028. Taiwan's entire economic weight, over 40% of its stock index in TSMC alone, sits inside one fabrication model whose neutrality is its actual product. Europe's role is one company in one town of 48,000 people. Japan's role is hidden inside the periphery: 88% of coaters, 53% of wafers, 50% of photoresists, 95% of packaging film, 90% of T-glass. Korea's is HBM. Each switch has been pulled at least once in the last four years, and the record of what happened when it was pulled is the strongest evidence this is not theoretical.

Is any of it producing value?#

It is fair to ask, at the end of a five-part anatomy of industrial concentration, whether the thing being concentrated actually works. The productivity evidence is unusually good for a technology this young, and worth listing because it comes from controlled experiments, not vendor decks. A randomized field experiment on GitHub Copilot found measurable speed improvements for developers. A study of customer-support workers found generative AI lifted their productivity substantially, with the largest gains for the least experienced. A randomized experiment on professional writing tasks found similar assistive gains. The Harvard-BCG field experiment found consultants improved significantly on tasks within AI's capability frontier, and degraded when they trusted it outside that frontier, which is the honest shape of the technology. AlphaFold predicted structures for roughly 200 million proteins, a contribution to science that nothing before it compares to in scale. In healthcare, 85% of surveyed leaders were exploring or had adopted generative AI. Enterprise usage grew about 8x in a year at OpenAI alone, Google reports 75% of new code is AI-generated, and the St. Louis Fed tracks AI contributing about 0.9 percentage points to quarterly US GDP growth.

Two epistemics notes belong here, because a series about who controls information infrastructure should apply the same skepticism to itself. The US government's own export-control justification includes a warning that advanced AI can lower barriers to cyber and CBRN capability, which is the stated reason for the gates, and NATO's framing of AI as a source of cognitive edge tells you the same thing from the other direction: the gates are not only commercial. And the AI research literature contains its own quality-control failures; one viral AI-materials-discovery paper was formally disavowed by MIT, which stated it had no confidence in the findings and requested its withdrawal. Extraordinary claims in this field deserve the same scrutiny as the claims in any other, including the ones on my own blog.

The economics of a chokepoint world#

The macro picture is a re-pricing in progress. AI is a capital-expenditure story, which is why the infrastructure makers have out-valued the application makers, and why a gigawatt-scale data center runs about $38 billion. The buildout is driving national economic statistics now, not just company earnings. Advanced packaging, the layer most people cannot name, is heading toward an $80 billion market by 2033. Korea and Taiwan rode exactly these layers past India in market capitalization. The OECD's economic security work now treats semiconductor value chains as a matter of national security on par with energy, which is the formal admission that the pipeline this series describes has become statecraft.

And the distribution question is not only about countries. The Oxford and OECD research on who owns versus rents compute, and the ITU projection of 68 gigawatts by 2027 and 327 by 2030, together describe a future in which the right to run AI is a licensed privilege for most of the world, granted through three clouds and two governments. That is not a conspiracy. It is an equilibrium, produced by the honest physics of the layers in parts 1 and 2. But equilibrium is not stability, and it is worth being precise about the difference.

What is stable, what is not#

The layers of the stack are stable in the sense that no challenger can dislodge any of them this year. They are unstable in a longer sense, and it is the same instability in every layer:

• Every monopoly has a challenger with a state behind it: the hyperscalers against NVIDIA, CXMT and the Korean trio in memory, SMIC and the DUV/EUV programs against TSMC-plus-ASML, Japan's and others' rare-earth substitution programs against Chinese refining
• Every chokepoint has been used at least once, which changes their character permanently: once the US showed licenses can gate chips, and China showed minerals can answer, every buyer on earth started paying the "sovereignty premium", paying more for redundant, second-source, geographically safer supply
• The demand curve does not care about any of this: training compute grows 4.5x a year, and 327 gigawatts of projected AI data center demand by 2030 does not care who holds which monopoly, it will simply find whichever path exists

The uncomfortable conclusion of the series is that the AI revolution, for all its software mystique, is physically a story about who owns seven or eight narrow industrial chokepoints, and the geopolitics of this decade is just those chokepoints being tested. The current arrangement resembles the oil order of the twentieth century, except that the "oil" is refined in one country, printed by one machine, packaged in one island, and sold through three companies. Oil had OPEC and tankers. This has ASML's shipping schedule and the Taiwan Strait.

What a country that owns none of it can do#

For the countries outside the walls, and India in particular as the case this series has followed, the honest option set is narrower than the headlines suggest, and worth stating without either despair or hype:

• Application layers are open to everyone, and they are not trivial. 67% of Indian AI startups there, building on the best models available, is a rational strategy given the constraint, and the productivity evidence says the value at the application layer is real. Swiggy inside ChatGPT, a Gemini agent in a Mahindra car, a bank running in-country Claude: these are real businesses on the pipeline's output.
• Services and talent compound. Nearly 20% of the world's chip design workforce being Indian, with the Indian government's EDA access program and the SHAKTI tapeouts as seeds, is the one asset that grows regardless of who owns the fabs. The 1984 story's lesson is not that India failed, it is that eight lost years can cost four decades. Talent is the asset that survived the fire.
• Negotiate from demand. 16% of the world's gen-AI downloads and a 10-12% share of the global AI-model user base is leverage of a kind: the buyers of 600 million downloads get localization (Gemini in Indian regions, Claude inference in India via Bedrock) that smaller markets do not get. That is leverage, and it is being used.
• Watch the two-decade clock. The Tata fab at 90nm by 2028, Semicon 2.0's packaging-first pivot, the 12 sovereign model teams: none of these make India a pipeline owner this decade. All of them change where India stands in 2040. Semiconductor strategy is measured in the timescale of the nodes it chases.

And a promise of method to close. Every number across these five parts came from primary documents: company filings, official releases, tracked datasets, and peer-reviewed studies, cross-checked where sources overlapped, with the celebrated claims (and my own enthusiasm for a good chart) discounted the same as the negative ones. The pipeline is the most documented industrial concentration in history, which is why writing it down matters: the map exists, it is verifiable, and more people should be looking at it than the ones who own it.

The next series goes downstream of the infrastructure: what happens on top of all this, the adoption patterns, the labor story, the data being produced by the people the models will affect most, and the application layer where most of the world actually lives. The infrastructure decides who can build. What gets built is the next question.

The full series#

• Part 1: Dirt, Sand and the Software That Draws Chips
• Part 2: One Machine and One Island
• Part 3: The Buyers and the Export Control Era
• Part 4: The Geography of Compute and the India Lens
• Part 5: Nobody Owns the Stack, Everyone Owns a Switch (you are here)