The Industrial Basis of AI Power
It has been observed that at the level of industrial production, having money and liquidity is indeed often life-saving, but not always the decisive variable. “It’s the only thing you can’t buy. I mean, I can buy anything I want, basically, but I can’t buy time.” Warren Buffett told Forbes, and he was so right. Because the delivery of the final product can be affected by other, perhaps boring and predictable, parameters that prevent a product from being delivered. A good example is Cisco’s story in the not-so-distant 2022. What they were missing was not some “rare” high-end subsystem that couldn’t be substituted. There were ready boards, there were orders, and there was demand that had already been locked in. What wasn’t there, at the right moment, were the simple parts that close the unit, like power supplies, small components, elements that under normal conditions you don’t even treat as “strategic.” And suddenly those ‘non-strategic’ parts produced a backlog “well over $15 billion” and the difference between “we have a product” and “we can deliver it.” (investor.cisco.com)
The lucky ones who read my previous article (and didn’t leave it to “mature” in some tab) remember that I mapped the geopolitics of AI across three levels. The upstream concerns the material inputs that support the hardware chain. The industrial level concerns converting that capability into real units that are produced, tested, and shipped. The downstream concerns whether those units can become operational power at scale, with electricity, installation, operation, and compliance. Here we will try to describe on the second level, because that is where “capacity” becomes a conversion chain, queues, and access rules. To research that we should take a good look at what Capacity means, what the bottlenecks are, how the power of concentration and how export controls operate actually.
This framing is not an “academic luxury” of a theorist, I believe it is becoming increasingly clear that Layer II is rising to the top of the international agenda, not only in G7 rooms but also in spaces where Global South states are trying to negotiate position within the new technological division of labor. At the AI Impact Summit 2026 in New Delhi, the summit closed with the adoption of the New Delhi Declaration on AI Impact and with an explicit tone of “AI as economic and strategic power,” not only safety. (mea.gov.in) At the same time, the political signal was accompanied by industrial language: major commitments for AI infrastructure and data centers as a framework for alliances and acceleration. (reuters.com)
Why “Capacity” looks more like a pipeline with multiple clocks
The word “capacity” here is, on the one hand, clear as a concept, but on the other it can become slippery if it is used as if it were a single number or a single scalar—X units per month, Y wafers per quarter, Z accelerators shipped. In Layer II it doesn’t behave like a number; it behaves like a pipeline with multiple clocks, not swiss made actually and rarely in sync. Allocation, wafer starts, packaging slots, HBM pairing, test time, burn-in, qualification, logistics, support, to name some of the clocks, each stage has its own cadence, its own constraints, its own failure modes, and those cadences rarely compress (or expand) together just because the market wants them to.
That’s why it’s so easy to misread what’s happening. A line can be “at capacity” and deliveries still don’t show up, because qualification slips, test racks become the bottleneck, or integration hits a calendar lock you can’t brute-force your way through. Strategically, the only capacity that matters is the one that survives end-to-end: units that pass, ship, and can be sustained—repeatably, as a system, not as a one-off surge.
Bottlenecks migrate and then migrate again across three major corners of the chain
That is exactly where the conversion chain comes in, starting from fabrication, continuing to advanced packaging that turns the wafer into a real accelerator, through the integration and qualification of memory (HBM), through test and burn-in that separates “paper output” from real output, and ending in delivery and support. Each stage functions like a gate. If one gate tightens and flow shrinks, the rest of the chain will not save you. Whatever procurement says, what counts at the end is “how many units passed, came out, and were delivered.” That is also why Layer II is geopolitical in the most mundane way as bottlenecks are not stable, they move. Today it might be packaging, tomorrow HBM, the day after test. And at the same time, access rules can make an otherwise realistic delivery “conditional.” To understand where power sits, you have to see where the queue sits.
At the industrial level it is natural to look for “the” bottleneck, the one lever that will decide everything. Historically, that is not irrational. In 1944, for example, the fuel constraint functioned as a single-piece bottleneck for the German war machine, when attacks on oil infrastructure led to a collapse in POL production and immediate operational consequences. (maxwell.af.mil)
In AI industrial layer, however, the pattern we capture is not a bottleneck that “sits” permanently somewhere and is more or less well defined. It is a bottleneck that moves. And if that sounds like an abstract schema, in practice it is the way the players themselves talk when they describe capacity planning, constraints, and pace: on one side, foundries talk about pressure from AI demand; on the other, vendors and their customers talk about queues and constraints in packaging and the back-end. (investor.tsmc.com)
So far, three constraints appear periodically.
Time-to-yield: the maturation time that even Mr. Warren Buffet cannot buy, because increasing production and yield has its own rhythm and follows cycles of weeks-to-quarters, as the industry itself describes it. (semiconductors.org) And when AI demand presses leading-edge processes, utilization rises and the margin for error shrinks. (investor.tsmc.com)
Packaging + HBM: where “more wafers” does not automatically mean “more compute,” because conversion into a final product gets stuck on conversion throughput and queues. One way to put it more bluntly is what Jensen Huang said: “packaging has remained a bottleneck due to capacity constraints.” (reuters.com) And next to packaging sits HBM, which often “locks” by calendar year: sold-out/near sold-out signals from memory suppliers show exactly that gate. (reuters.com)
Test & burn-in: the silent governor, because a unit is not real supply until it passes validation without creating a new queue. Test vendors themselves describe this as “test time” and “thermal control” in HPC/AI. (advantest.com) And when a company like Nvidia openly invests in cutting testing time dramatically, it is hard to read it as anything other than bottleneck behavior. (reuters.com).
Concentration is Power
Layer II is not a description of a production line in a mature market, but production inside a landscape of concentration, where a few nodes carry disproportionate weight and substitutions are slow. Consider that three companies hold the lion’s share of global design software; we are not really talking about a “free market with choices.” Here we are talking about EDA (Electronic Design Automation), meaning the software through which chips are designed, verified, and “locked” (sign-off) before tape-out and fabrication. And when workflows, licenses, and support are concentrated, you do not change ecosystems “because you found a better offer.” According to TrendForce, in 2024 the shares for Synopsys, Cadence, and Siemens EDA are roughly 31%, 30%, and 13% respectively. (trendforce.com)
The same logic applies to the critical steps of the conversion chain we described. As technology becomes more complex, the cost of switching suppliers rises and qualification time grows. This is geopolitical not because someone “flips a switch,” but because concentration makes every bottleneck stickier and translates into a very practical conclusion: when a node tightens, you don’t just “switch.” You wait, you adapt design, you join the queue, because you simply don’t have clean alternatives. Unless you are Tesla and, during the semiconductor crisis, you manage to do tweaks and substitutions and move faster than everyone else. (reuters.com) But how many have that scale and boldness?
When this concentration becomes an interstate issue, the response does not come only from the market but also from alliance structures. Pax Silica is presented by the U.S. State Department as a “flagship effort on AI and supply chain security,” meaning an attempt to turn dependencies into institutionally organized resilience. (state.gov) In geopolitical terms, we are witnessing an analogy of Halford Mackinder’s 'Heartland Theory.' It is not a 1:1 analogy but i believe it illustrates that just as the 'World-Island' was once the pivot of history, the 'Industrial Heartland' of EDA software and foundries has become the new pivot, who rules this industrial core doesn't just manage a market they are in command of the global compute-stack.
Are export controls friction by design?
In Layer II, export controls rarely resemble the cinematic “switch” that cuts the power or taps the natural gas pipeline. More often they resemble friction in the calendar, precisely where productive capability must become deliverable product. I am not saying this as a supply-chain specialist. I am saying it as an IR scholar reading public decisions, corporate statements, and the way these translate into pace and predictability.
BIS’s move on January 13, 2026 is indicative precisely because it is “mundane”: a change in licensing posture for exports of certain advanced chips to China, with a case-by-case review framework under conditions. (bis.gov) A few days later, the same change appears as a regulatory fact in the Federal Register: Revision to License Review Policy for Advanced Computing Commodities (January 15, 2026), FR Doc. 2026-00789, Docket No. 260112-0028, 91 FR 1684, RIN 0694-AK43. (federalregister.gov)
Two mechanisms produce leverage without requiring an outright ban. Noone is suprised to see firms planning defensively when licensing becomes less predictable. And as any procurement manager knows, acquisition is not a one-off purchase. It includes sustainment—whether the system can be supported, repaired, and upgraded. Once spares and support move into a cautious, case-by-case posture, refresh cycles slow. In Layer II, that loss of tempo is often enough.
But what if export controls ultimately function like a “gym”? Instead of cutting capability, do they push the restricted actor to become more efficient and accelerate domestic alternatives? The China case makes it hard to ignore, it is just 10 months now since Jensen Huang called the U.S. restrictions a “failure,” at Reuters saying they lit a fire under China to intensify its push toward a more autonomous ecosystem. The risk is real over a long horizon, but it does not cancel the Layer II leverage. Adaptation does not eliminate the bottleneck; it changes how you live inside it. And for much of the game, power is decided by the pace and predictability of the pipeline, where queues and qualification gates operate in quarters. In the language of strategist John Boyd, Layer II is about tempo. By controlling the friction of the next move, you force the adversary to operate on an obsolete map of the battlefield, effectively winning by ensuring they can never quite catch up to the current loop of innovation. I know someone might say that quoting a general and using a word like adversary is maybe too much, but are we really operating not in an adversarial framework?
The same pattern is visible in equipment. ASML, on the updated U.S. restrictions of December 2024, speaks in a language that is not “commentary on geopolitics,” but effective dates and compliance deadlines. (asml.com) That is how rules become part of the industrial calendar, not merely a framework around it. Historically, this way of exercising power is not new. COCOM in the Cold War and the subsequent Wassenaar logic show how controls on technology diffusion work through lists, technical categories, and national bureaucracies, meaning through processes that produce time and uncertainty as political outcomes. (everycrsreport.com)
Alongside export controls, a quieter layer of rules is also being built through reporting and comparability. In February 2025, the OECD launched the Hiroshima AI Process (HAIP) Reporting Framework as a monitoring/reporting mechanism linked to the Hiroshima AI Code of Conduct. (oecd.org) This does not replace the hard leverage of licensing. It does, however, add visibility and institutional normality.
And here an open question remains, without opening a separate European section: can the Brussels Effect acquire geopolitical value in Layer II, not because it increases throughput, but because it makes rules become de facto global defaults? (academic.oup.com) If this translates into certification, insurability/financing, and contractual support terms around advanced compute, then “power” would also be rule-setter advantage. It will only be visible, however, if market access is central enough to impose harmonization, not merely cost.
Author’s commentary:
I didn’t explicitly say this in the article, but the heart of the series is about mapping AI as a national capability. I’m trying to map the layers in order to examine how government capacity, corporate execution, and human talent combine through a grand-strategy lens. To me, a single breakthrough or a “heroic” training run matters less than whether a country can keep upgrading its stack when the pipeline tightens. In practice that cashes out in allocations, licensing posture, service access, and refresh cycles. I’m trying to pin down those mechanisms so we don’t get pulled around by headlines and can see where the next bottleneck is likely to land.
Important notice: This piece is part of a broader effort to map the infrastructural and governance layers of AI geopolitics. A separate academic article with a narrower research question and formal conceptual framework is under development. Any comment is welcome on the layers foundation presented here. LLM assistance was limited to light copyediting (clarity/grammar) and image iteration. Research, argument structure, and source verification were done by the author.
References:
Cisco Systems, Inc. (2022, May 18). Cisco reports third quarter earnings (reference to product backlog “well over $15 billion”). (investor.cisco.com)
Ministry of External Affairs, Government of India. (2026, February 21). AI Impact Summit 2026 concludes with adoption of New Delhi Declaration on AI Impact. (mea.gov.in)
Ministry of External Affairs, Government of India. (2026, February 21). AI Impact Summit Declaration, New Delhi (February 18–19, 2026). (mea.gov.in)
Reuters. (2026, February 19). Tech majors commit billions of dollars to India at AI summit. (reuters.com)
Air University / Maxwell AFB. (2019, June 26). WWII Allied “Oil Plan” devastates German POL production. (maxwell.af.mil)
TSMC. (2024, Q2). 2Q24 Earnings Conference Transcript (PDF). (investor.tsmc.com)
Semiconductor Industry Association. (2021, February 26). Chipmakers are ramping up production… Here’s why that takes time. (semiconductors.org)
TSMC. (2024, Q3). 3Q24 Earnings Conference Transcript (PDF). (investor.tsmc.com)
Reuters. (2025, January 16). Nvidia CEO says its advanced packaging technology needs are changing. (reuters.com)
Reuters. (2024, May 2). Nvidia supplier SK Hynix says HBM chips almost sold out for 2025. (reuters.com)
Reuters. (2026, February 12). Samsung says it has shipped HBM4 chips to customers. (reuters.com)
Advantest. (2024, November 28). Characteristics and Needs of HPC/AI Device Test (IR Technical Briefing). (advantest.com)
SEMI. (2025, December 15). Global semiconductor equipment sales projected… (reference to test equipment +48.1% for 2025). (semi.org)
Reuters. (2025, November 19). Nvidia, Menlo Micro collaboration speeds up AI chip testing. (reuters.com)
TrendForce. (2025, June 2). China revenue at risk as U.S. curbs slam EDA giants… (EDA shares 2024: 31/30/13). (trendforce.com)
Reuters. (2022, January 4). Explainer: How Tesla weathered global supply chain issues that knocked rivals. (reuters.com)
U.S. Department of State. (n.d.). Pax Silica. (state.gov)
Mackinder, H. J. (1919). Democratic Ideals and Reality: A Study in the Politics of Reconstruction. Henry Holt and Company.
Bureau of Industry and Security (BIS). (2026, January 13). Department of Commerce revises license review policy for semiconductors exported to China. (bis.gov)
Federal Register. (2026, January 15). Revision to License Review Policy for Advanced Computing Commodities (FR Doc. 2026-00789; Docket No. 260112-0028; 91 FR 1684; RIN 0694-AK43). (federalregister.gov)
Boyd, J. R. (1995). The Essence of Winning and Losing. Patterns of Conflict (manuscripts).
ASML. (2024, December 2). ASML statement on updated US export restrictions. (asml.com)
Congressional Research Service. (2006, September 29). Military Technology and Conventional Weapons Export Controls: The Wassenaar Arrangement (RS20517). (everycrsreport.com)
Wassenaar Arrangement. (n.d.). Genesis of the Wassenaar Arrangement. (wassenaar.org)
OECD. (2025, February 7). Launch of the Hiroshima AI Process (HAIP) Reporting Framework. (oecd.org)
Bradford, A. (2019). The Brussels Effect: How the European Union Rules the World. Oxford University Press. (academic.oup.com)
Yole Group. (2025, July 29). Wafer Fab Equipment (WFE) market to hit $184 billion by 2030… (Press release). https://www.yolegroup.com/press-release/wafer-fab-equipment-wfe-market-to-hit-184-billion-by-2030-for-equipment-and-services-driven-by-specialized-segment-growth-and-global-manufacturing-shifts/