CEO Letter: The Wrong Factory

America is building again. You see it in the cranes over the cornfields in Iowa, the concrete pads rising from flattened scrubland outside Phoenix, the substation upgrades running through the Appalachian grid, new arteries cut into old tissue. Hundreds of billions of dollars are flowing into the ground , copper wire, fibre optic cable, cooling systems, server racks stacked to the ceiling of buildings the size of aircraft hangars.

The commentators who tell you America doesn't make things anymore are looking at the wrong thing. They're looking for the old factory , the one with a thousand workers on a floor, punch clocks and union halls and a parking lot that is filled at six in the morning. That factory is gone. But something is being built in its place. The question worth asking, the question that keeps getting swallowed by the technology narrative, is whether what's being built serves the country that's building it.

Data centres are the reindustrialisation of the United States. The investment is real, the physical infrastructure is real. Microsoft committed $80 billion to data centre construction in fiscal year 2025 alone. Combined hyperscaler capital expenditure across Microsoft, Amazon, Alphabet, and Meta went from $69 billion in 2019 to an estimated $670 billion in 2026 , nearly a 10-fold increase in seven years. McKinsey estimates that global data centre infrastructure investment will total $7 trillion by 2030. By the measure of capital formation, the United States is in the middle of a genuine industrial renaissance.

But the old industrial towns are still hollowed out. The parking lot still doesn't fill.

THE ECONOMICS OF LABOUR-LIGHT CONSTRUCTION

This is the contradiction at the centre of what is being built. The data centre construction boom employs construction workers , electricians, ironworkers, concrete pourers , while the work lasts. It employs a thin layer of highly skilled technicians to maintain the facilities once they're running. A data centre the size of a small city might run permanently with a staff of sixty people. The factory it replaced employed thousands.

The new industrialisation is capital-intensive and labour-light by design. A semiconductor fab requires precision clean rooms and specialist operators. A data centre requires concrete, cable trays, and cooling , and then mostly automated systems and a small maintenance crew. The capital formation number is real. The jobs multiplier is not what the headline suggests.

This would be tolerable , perhaps even net positive , if the productivity gains it enables were broadly distributed. Technology has always disrupted labour markets, and the historical argument for tolerating short-term displacement is that the wealth created eventually finds its way into wages, living standards, and opportunity. The spinning jenny put hand-weavers out of work and eventually clothed the world at a price ordinary people could afford.

The timeline of AI is not the timeline of the spinning jenny.

FROM FREE CASH FLOW TO BORROWED MONEY

Until recently, hyperscalers funded their capital programmes from their own operating cash flows. That era has ended.

For the first two decades of cloud infrastructure buildout , roughly 2005 to 2023 , Amazon, Google, Microsoft, and Meta deployed capital at a pace their operating businesses could sustain. AWS revenue funded AWS infrastructure. Google Search funded Google Cloud. The virtuous cycle of cloud economics , spend capital, generate recurring revenue, spend more capital , was self-reinforcing and internally financed.

The AI capex surge has broken that model.

Amazon's free cash flow trajectory tells the story most starkly. In 2024, Amazon generated $38.2 billion in free cash flow , healthy, growing, well above its capital expenditure. In 2025, Amazon's free cash flow collapsed to $11.2 billion , a 70% decline in a single year, as capital expenditure for the year totalled ~$132 billion — far exceeding the $97 billion Morgan Stanley had forecast in November ‘24. By Q2 2026, Amazon had turned FCF negative ($-7.6 billion) on a quarterly basis for the first time. CNBC, reporting in February 2026, noted analyst commentary: "If you're going to pour all this money into AI, it's going to reduce your free cash flow."

Epoch AI, in a June 2026 analysis, projected that aggregate capital expenditure across Microsoft, Amazon, Alphabet, Meta, and Oracle would overtake their combined operating cash flow in Q3 2026. FactSet stated it plainly: "The unprecedented increase in AI investments has pushed [hyperscalers] toward other avenues of financing. Since FY24, free cash flow has trended downward for most players and is expected to decline further in FY26."

The "other avenues of financing" are debt markets, at a scale that would have been unthinkable five years ago.

Bank of America analyst Yuri Seliger documented the shift: the five largest hyperscalers raised roughly $140 billion in bonds across the entire 2020–2024 period. In 2025 alone, they issued $121 billion. In the first six months of 2026, they issued approximately $195 billion , more in six months than in all of 2025.

The landmark deals are worth noting individually. Meta issued $30 billion in senior notes in October 2025 , the largest bond offering since Pfizer's $31 billion in 2023. Oracle completed an $18 billion bond offering in September 2025, originally marketed at $15 billion; demand reached $88 billion, nearly five times oversubscribed, with proceeds explicitly earmarked for AI infrastructure. Amazon issued $25 billion in bonds in July 2026. Alphabet issued multi-currency bonds including a 100-year sterling tranche. Fortune headlined it as "Google, Meta, and Oracle are on a $1 trillion borrowing spree."

There is also a less visible dimension. Nikkei, reporting in July 2026, identified approximately $1.65 trillion in off-balance-sheet obligations , build-to-suit data centre leases, co-location commitments, GPU power purchase agreements , across America's five largest AI infrastructure companies. More than their combined on-balance-sheet debt, but not visible in the headline leverage ratios that most investors see.

The hyperscalers are not in financial distress. Their balance sheets remain strong. But the structural shift is significant: the build-out of AI infrastructure has moved from a capital allocation decision, how do we deploy our surplus cash most productively? To a financing decision. The industry that for two decades defined self-funded compounding growth is now borrowing at scale to fund an investment whose returns remain hypothetical.

THE DEPRECIATION PROBLEM

There is a number at the centre of every hyperscaler income statement that determines how profitable the AI investment appears to be. It is the depreciation line. And it has been quietly restructured.

Between 2023 and 2024, every major hyperscaler extended the assumed useful life of its server and networking hardware , a change that, collectively, reduced their annual depreciation expense by approximately $18 billion. The change is not controversial in accounting terms; useful life assumptions are legitimate management judgements. But the timing and direction of the change create a striking circularity.

Google, reporting in February 2023, disclosed that effective January 2023 it was extending server useful lives from four years to six years , a change that reduced annual depreciation expense by approximately $3.4 billion in FY2023 alone. Meta made the same change within days of Google's announcement, extending from four years to five, and then to five-and-a-half years by January 2025, generating a $2.92 billion reduction in depreciation expense and a $1.00 increase in diluted EPS for full-year 2025. Microsoft and Amazon normalised their own depreciation assumptions in the same direction, at approximately six years.

The accounting rationale: hardware lasts longer than the original assumptions suggested. The economic reality: GPUs depreciate faster than the accounting suggests.

An H100 GPU purchased in 2022 at $30,000 to $40,000 was worth approximately $12,000 to $15,000 on the secondary market by early 2026 , a 60-73% value destruction in three years. The Nvidia generational cadence runs at approximately 18-month intervals: A100 (2020), H100 (2022), H200 (late 2023), B200/Blackwell (2025), B300 (2026). Each new generation renders prior-generation hardware significantly less competitive for frontier training workloads. The market price of a used H100 reflects this. The straight-line depreciation schedule does not.

AltStreet, in January 2026, put the divergence precisely: actual GPU depreciation is frontloaded at approximately 35% in Year 1 , reflecting the rapid obsolescence when new generations arrive , versus the accounting assumption of approximately 17% per year on the six-year straight-line schedule. The hyperscalers extended depreciation lives just before the capex curve went vertical, precisely when investors were most scrutinising their return on invested capital. The result is that reported earnings look better than economic reality warrants, at exactly the moment when the scale of the bet is largest.

This matters beyond accounting tidiness. If the H100s purchased in 2022 and 2023 are economically obsolete by 2025–2026 , if the models training on them have moved to H200s and B200s and the inference workloads are increasingly running on smaller, more efficient architectures , then the depreciation charge being reported is materially understating the economic cost of the capital deployed. The earnings that justify the continued investment may be structurally overstated. And the capex cycle requires each new generation of hardware to be purchased, at scale, before the previous generation has been written off.

THE DEMAND QUESTION

In June 2024, Jim Covello, Head of Global Equity Research at Goldman Sachs, published a report titled "Gen AI: Too Much Spend, Too Little Benefit?" It was notable not for the caution, there is no shortage of AI scepticism, but for its specificity. Covello estimated that tech giants and beyond were set to spend over $1 trillion on AI capital expenditure in coming years. He noted that only 6.1% of US companies were actually using AI for business functions, per the US Census Bureau’s Business Trends and Outlook Survey in Q4 2024.. He argued that AI was "exceptionally expensive" relative to the problems it currently solves, and that the cost of the technology needed to fall by orders of magnitude before it could justify the investment required to build the applications that would generate the revenue that would pay back the infrastructure.

The revenue gap is stark. Total generative AI market revenues in 2024 were approximately $30 billion. Morgan Stanley estimates $2.9 trillion in global data centre construction planned between 2025 and 2028. That is a 97:1 ratio of planned supply to current revenue generation. Goldman Sachs projects that global data centre power demand will rise 165% by 2030 relative to 2023 baseline levels, from roughly 55 gigawatts to 122 gigawatts of capacity. The IEA projects data centre electricity consumption to double from 415 TWh in 2024 to 945 TWh by 2030. More than the entire current electricity consumption of Japan.

The counterargument is that demand follows supply, as it did with cloud computing: the infrastructure was built ahead of the applications, and the applications eventually materialised and exceeded the original projections. This is possible. The structural productivity gains from AI are genuine and in some domains already measurable. The Goldman analysts who disagreed with Covello within the same report argued that the potential returns from the current AI capex cycle are "not significantly different from prior tech investment cycles."

But the demand case for AI infrastructure rests on a specific hypothesis: that AI will generate sufficient commercial value in productivity savings, new product categories, cost displacement to justify infrastructure investment at $7 trillion over six years. That hypothesis is not wrong. It is simply unproven at the scale required. Mark Cuban, in July 2026, put it more bluntly: "A lot of data centers are going to be turned into pickleball courts." He drew the explicit parallel to the 1990s fibre-optic overbuild , $500 billion invested globally in fibre infrastructure, companies like WorldCom, Global Crossing, and Qwest imploding, most fibre running dark for years. "Demand for bandwidth didn't materialise as quickly as expected," as one analyst put it. The infrastructure eventually found its purpose , the fibre that ran dark through the 2000s became the backbone of the internet economy of the 2010s. But the companies that built it mostly didn't survive to benefit.

COMPETITION WITH CHINA CREATING DOWNWARD PRICING PRESSURE, RAISING THE PRODUCTIVITY BAR TO CLEAR THE ROI HURDLE RATE

There is a pattern to how China enters a market it intends to own. The prices arrive first , sharply, almost offensively below what anyone thought was commercially possible. Then the Western competitors begin their slow retreat: margins compress, investment cases fall apart, factories close. Then, quietly, the dependency installs itself. By the time the strategic intent becomes legible, the switching costs are prohibitive.

This is not a theory about AI. It is a documented industrial playbook that has already run to completion in solar panels, steel and telecommunications equipment. It is now running in AI inference. The question is not whether to recognise the pattern. It is whether recognition comes early enough to matter.

The solar case is the template. In 2010, a solar module (panel) cost roughly $1.48 per watt to produce. By 2023, China's production cost was $0.15 per watt , against $0.40 per watt in the United States and $0.30 per watt in Europe. That is a 60 to 75 percent structural cost advantage, not a temporary promotion. The route to that advantage was explicit: China Development Bank and the Export-Import Bank of China extended over $30 billion in subsidised credit to Chinese solar manufacturers between 2010 and 2012, enabling years of below-cost pricing that the global market could not match. BP Solar closed in 2011. Q-Cells, Germany's largest panel manufacturer, went bankrupt in 2012 and was acquired by a Korean conglomerate. SolarWorld followed in 2017, then again in 2019. China now produces approximately 80 percent of the world's solar panels. The supply chain is not contested. It is settled.

The AI chapter opened with a single event: DeepSeek V2, released in May 2024 at one yuan per million input tokens and two yuan per million output tokens. That was approximately one-seventieth the price of GPT-4 Turbo at the time. On the day of release, the pricing structure of China's entire large language model industry collapsed within a week , Alibaba cut its model prices by 97 percent within days. Baidu and Tencent followed. A new floor had been established, and it was a floor the Western incumbents had not priced for.

DeepSeek R1, released in January 2025, made the international dimension of this impossible to ignore. R1 matched OpenAI o1 on mathematical olympiad benchmarks , 79.8 percent against 79.2 percent on AIME 2024, 97.3 percent against 96.4 percent on MATH-500, ahead on competitive coding. Its base model, DeepSeek V3, was trained for approximately $5.6–6 million using around 2,000 H800 GPUs — a figure that refers specifically to the final training run compute cost, excluding prior R&D, data preparation, and infrastructure investment. GPT-4's training is estimated at $80 to $100 million. The R1 API today prices at $0.55 per million input tokens and $2.19 per million output tokens. OpenAI o1 prices at $15 and $60 respectively. That is a 27-fold price differential at equivalent or superior performance on the benchmarks that matter most to enterprise developers.

Alibaba's Qwen models sit at 6 to 12 times below GPT-4o and Claude Opus equivalents at comparable capability tiers. ByteDance's Doubao-1.5-Pro was described by US News on its launch as "undercutting rivals by 99 percent." The cheapest ByteDance inference is now available below $0.04 per million input tokens.

The question of whether this pricing reflects genuine engineering efficiency or deliberate strategic pricing is largely the wrong question. The answer is both, and the distinction is immaterial to the outcome. DeepSeek's Mixture-of-Experts architecture , activating only 37 billion of its 671 billion parameters per inference , genuinely reduces compute costs. The Multi-head Latent Attention mechanism genuinely reduces memory requirements. These are real innovations, not accounting fictions. But genuine efficiency creates the capability to price low. It does not compel the decision to price at 1/70th of a competitor. That decision is strategic. MERICS named it directly in June 2026: "wide dispersion, cheap tokens" , a policy of blanketing the global market with near-free AI access to capture developer mindshare and create platform dependency, with profit explicitly secondary to market position and geopolitical influence.

The mechanism is the same across all of these cases and it runs in four phases. In phase one, state-backed capital subsidises pricing below the Western cost of production. Western competitors' investment cases collapse. Factories close, research programmes are cut, talent disperses. In phase two, the volume acquired through loss-leading drives genuine scale economies , the learning curve kicks in and the cost advantage becomes real rather than subsidised. In phase three, customers build on the cheap infrastructure. Switching costs accumulate. A developer who has built agents on Qwen APIs, whose fine-tuned models run on Alibaba Cloud, whose data flows through Chinese infrastructure , that developer faces years of migration work to move. In phase four, the platform controls the stack. The pricing can rise. The relationship is structural.

The open-source angle in AI is a refinement of the playbook that did not exist in solar or steel. DeepSeek and Qwen release model weights publicly. You cannot ban a GitHub repository with an export control order. The model propagates globally, runs on local hardware, and establishes cognitive dependency without requiring infrastructure control. The API inference revenue and data flows that accrue to Chinese platforms come from the developers who find the hosted version convenient , which, at $0.04 per million tokens, most of them do.

The GPU export controls that the United States has imposed since 2022 , restricting access to Nvidia's highest-capability chips , were premised on the theory that compute scarcity would slow Chinese AI development. DeepSeek's training cost of $6 million against GPT-4's $100 million is the empirical response to that premise. The chokepoint that was supposed to hold did not hold. The engineering work to route around hardware constraints is now the source of the cost advantage being used to price Western competitors out of global developer markets.

There is a version of this that ends with the AI infrastructure of the global economy running primarily on Chinese models and Chinese clouds , not through conquest, but through convenience. The price was right. The performance was equivalent. The switching costs accumulated quietly. That is how it ended in solar. That is how it is proceeding in steel. The current investment in data centre capacity that this essay opened with , $670 billion across the hyperscalers in 2026, debt-financed, premised on demand projections that have not yet materialised , is the Western response to a competitive threat that undercuts the entire financial model on which that investment is premised. You cannot depreciate a $100 million data centre over six years when your primary competitor trained at $6 million and charges $0.55 per million tokens.

China is not disrupting the AI industry. It is running a playbook it has already completed elsewhere. The factories that should worry about this are not the ones making steel.

AND THAT BRINGS US TO: THE RETIREMENT TRAP

There is a structural irony at the centre of this investment cycle that has received almost no attention.

The economic thesis propping up the current data centre boom is a bet that artificial intelligence will generate sufficient commercial productivity to justify the capital being deployed. That bet is embedded in the equity valuations of the hyperscalers , and through those valuations, in the pension funds, 401(k) accounts, and sovereign wealth vehicles that hold their shares. Fifty million American workers have retirement savings that are, in some measure, exposed to the assumption that AI demand will materialise at the scale implied by current valuations.

The trap is this: if the bet loses. If AI demand fails to materialise at the scale required. If the demand is there but the value has eroded… the capital destruction falls on the retirement accounts of precisely the Americans who were displaced by the last wave of automation. The factory worker who lost their job in the 90s and the middle class knowledge worker who lost their job in 2025, through their pension fund, financing the data centre that will decide whether their retirement is comfortable or precarious.

And if the bet wins , if AI does what its advocates claim, if it becomes the general-purpose technology that electricity was, if it does reach into every sector of the economy and reshape how work is done , the displacement accelerates. The productivity gains accrue to capital. The jobs that remain are those that AI cannot yet do. A shrinking list, compiled by the same technology that is eliminating the others.

This is the trap. The country needs the bet to win to preserve the asset values holding up retirement security. But winning the bet means deploying the technology at scale, and deploying it at scale means accelerating the structural unemployment that is already corroding the middle class. The retirement system is being used to finance the instrument of its own beneficiaries' displacement.

There is a version of this story that ends well. It requires redistribution mechanisms sophisticated enough to move productivity gains from capital to labour, education and retraining infrastructure built at the scale of the disruption rather than at the scale of the political appetite, and a government willing to treat the transition as the national emergency that it is rather than as the innovation success story that the technology industry prefers to tell.

America is building again. The cranes are real, the debt is real, the depreciation schedules have been carefully extended, and the demand is hypothetical.

The factory is back. Reindustrialisation is here… it just looks difference this time.

TTFN not forever.
— Nick