The Assumption Is Breaking

A structural reckoning is underway in enterprise technology — one that most CFOs, boards, and CIOs have not yet fully priced into their infrastructure strategies.

The AI arms race has changed the economics of compute infrastructure. Hardware that costs billions to acquire is economically obsolete in roughly three years — before it is even half-depreciated on a company’s books. Hyperscalers are spending hundreds of billions annually in a cycle that more closely resembles supermarket restocking than industrial investment. And the chip supply chain that powers this ecosystem is being cornered by entities with balance sheets no enterprise can match.

This is not a technology story. It is a capital markets story. And it has a direct parallel in recent financial history.

Six forcing functions are simultaneously ending the era of private enterprise hardware ownership:

1. Economic obsolescence — a 3-year economic asset life on hardware depreciated over 5–6 years

2. Supply foreclosure — hyperscalers and sovereign buyers cornering chip access at a scale no enterprise can match

3. Operational drag — the talent, energy, and management overhead of an accelerating hardware cycle

4. Architectural uncertainty — the dominant compute paradigm may itself be transitional

5. Accounting distortion — GAAP treatment masks true obsolescence velocity; a regulatory reckoning is forming

6. Geopolitical fracture — Taiwan concentration risk, the U.S.–China tech war, and sovereign AI demand reshaping the entire supply chain

This paper makes no recommendation about where any enterprise should place its infrastructure. It offers a diagnosis: the assumption that AI infrastructure investment is durable and productive in the way the accounting and the markets assume is breaking — and the enterprises that understand why will make fundamentally different decisions than those who don’t.

The SaaS Parallel

When “Safe” Revenue Wasn’t

To understand what is happening to AI infrastructure investment today, it helps to remember what happened to SaaS valuations yesterday.

For most of the 2010s, SaaS companies commanded premium multiples on a simple premise: annual recurring revenue was safe revenue. Subscription contracts, high switching costs, net revenue retention above 100% — the narrative was that once a customer was in, they stayed in, and the revenue compounded reliably.

The market priced that narrative generously. Revenue multiples of 15x, 20x, even 30x became common for high-growth SaaS businesses. Investors were not just paying for current earnings — they were paying for the assumed quality and durability of the revenue stream.

Then the assumptions changed. AI began commoditizing SaaS features at a rate the models had not accounted for. Point solutions protected by workflow lock-in suddenly faced competition from general-purpose AI tools that could replicate core functionality without a subscription. Churn risk, previously theoretical, became measurable. The “recurring” in ARR turned out to be more conditional than the multiples had implied.

The market repriced — not because the revenue disappeared, but because the quality of that revenue was lower than advertised. The safety assumption was wrong, and when the market recognized it, the correction was swift and significant.

What Research Affiliates CEO Chris Brightman documented in April 2026 is that the premise of durable, compounding infrastructure investment is structurally false in the AI era. The capex is not compounding. It is churning. And the churn cycle is accelerating.

The parallel is exact. SaaS ARR appeared safe until the competitive dynamics that made it safe changed. Cloud capex appeared productive until the innovation cycle that defined its useful life accelerated beyond what the accounting models assumed. In both cases, the market priced an assumption. In both cases, the assumption broke.

Six Forcing Functions

The Convergence Ending Private Hardware Ownership

The end of private enterprise hardware ownership is not a prediction. It is a convergence — six independent forcing functions arriving simultaneously, each sufficient on its own to shift the calculus, together creating a structural irreversibility.

1 — Economic Obsolescence

The Brightman analysis establishes the foundational data point: AI hardware has an economic life of approximately three years, while it is depreciated over five to six years on corporate income statements.

The proof is in the unit economics of Nvidia’s H100 GPUs. In year two of deployment, an H100 generated a return on investment of 137%. By year four, the same hardware was producing a negative ROI of 34%. The curve does not flatten — it accelerates downward.

This gap between accounting life and economic life has profound implications. Earnings are systematically overstated during the useful period of the asset. Write-downs are coming as hardware becomes economically obsolete before it clears the books.

2 — Supply Foreclosure

One number reframes the argument. Announcing the Terafab project on March 21, 2026, Elon Musk stated that the combined output of every advanced semiconductor foundry operating on Earth represents approximately 2% of the compute his companies alone will need. The entire current global semiconductor industry — TSMC, Samsung, Intel, all of it — is 2% of one company’s projected demand.

The data center buildout has already run into the supply wall. Of 12 gigawatts of planned US data center capacity, approximately 50% is delayed or canceled, 33% is under construction, and 17% remains uncertain. The cause is not chip scarcity alone — it is electrical equipment shortage: transformers, switchgear, and power infrastructure components. The constraint has moved upstream of the chips themselves.

ASML holds a near-absolute monopoly on extreme ultraviolet lithography — the EUV machines required to manufacture semiconductors at advanced nodes — producing approximately 50 systems per year globally. Every advanced AI chip flows through that single bottleneck. Tracking cumulative AI chip ownership through 2025, the global stock approaches 20 million units, with Google holding the dominant position across both TPUs and H100s. The enterprise market does not appear on this chart. It is structurally absent from the ownership class that will define AI compute capacity for the next cycle.

3 — Architectural Uncertainty

TSMC is a process innovator, not an architecture innovator. Their manufacturing excellence is not in question. But process leadership is not architecture leadership — and the most consequential players in the AI economy are determining architecture themselves, divergently, in parallel.

The pattern is unambiguous: every major player with the balance sheet to make the investment is actively routing around the Nvidia/TSMC-dependent GPU architecture. Any enterprise that has invested in private hardware ownership has locked capital into an architectural bet that the largest players have already decided they are not willing to make.

4 — Operational Drag

Managing competitive AI infrastructure requires continuous hardware evaluation, procurement, deployment, and retirement on an accelerating cycle. It requires energy infrastructure that scales with each generation, and a specialized talent base staying current with an architecture landscape changing faster than most organizations can hire.

For hyperscalers, this is their core business. For every other enterprise, it is overhead — distraction from the actual competitive work of building products, serving customers, and generating returns. Every dollar spent managing hardware is a dollar not spent building the actual moat.

5 — Accounting Distortion

AI hardware with a 3-year economic life is being depreciated over 5–6 years on corporate income statements. When an asset is depreciated more slowly than it loses economic value, the depreciation charge understates the true cost of maintaining the revenue that asset generates. Income appears higher than it would if the accounting reflected economic reality. Return on capital appears stronger. The business looks more profitable than it is.

At the scale of current annual AI capex, the earnings overstatement is not a rounding error. It is a structural feature of current financial reporting that investors, analysts, and boards are only beginning to scrutinize. The FASB has not yet moved on accelerated depreciation for AI infrastructure. But when hundreds of billions in annual capex produce shrinking margins and 3-year economic obsolescence, the SEC will eventually ask why income statements reflect 5–6 year depreciation. The regulatory reckoning is forming.

6 — Geopolitical Fracture

TSMC manufactures approximately 90% of the world’s advanced semiconductors — in Taiwan. A geopolitical event in the Taiwan Strait would not disrupt one company. It would simultaneously sever the advanced chip supply for every hyperscaler, every national AI program, every enterprise, and every defense system that depends on leading-edge silicon. This is not a supply chain risk in the conventional sense. It is a civilizational infrastructure dependency concentrated in the most geopolitically contested geography on earth.

The United States entered the sovereign AI tier at maximum scale with the Stargate Initiative — a $500 billion AI infrastructure commitment competing for the same chip supply, energy capacity, and construction resources as every other buyer. When a government declares AI infrastructure a national strategic priority, every other buyer in the queue moves down one position. India is emerging as a consequential new pole — over $10 billion in incentives, Tata Electronics entering chip manufacturing, and an explicit policy to reduce Taiwan-concentrated dependence. And DeepSeek’s early-2025 achievement — near-frontier performance at a fraction of assumed compute cost — was a direct geopolitical response to chip scarcity: adversarial constraint producing efficiency innovation the unconstrained model had no incentive to pursue.

The Number Changed

The Mechanism Is Running Faster Than the Models Assumed

This paper was first written in April 2026 around a single foundational number: roughly $650 billion in annual AI infrastructure spend across the major hyperscalers. That number anchored every argument — the obsolescence math, the supply foreclosure dynamic, the accounting distortion. It was the floor the diagnosis was built on.

Four weeks later, Morgan Stanley revised it.

A 24% upward revision in four weeks. The number didn’t drift — it accelerated. And the acceleration is itself the data point that matters most. The revision doesn’t weaken the diagnosis; it tightens it. The obsolescence gap between a 3-year economic life and a 5–6 year depreciation schedule doesn’t shrink when annual capex grows from $650B to $805B. It widens. The earnings overstatement grows in proportion to the investment. The accounting reckoning, when it comes, arrives at a larger number.

The Jevons Dynamic, Visible in Real Time

The hyperscalers aren’t spending more because the technology got more expensive. They’re spending more because demand for what the technology enables is expanding faster than the prior models projected. As AI reduces the cost of cognition, the efficiency dividend is consumed by an explosion of new uses before it can accumulate — deployment accelerates demand for compute, which accelerates investment, which accelerates deployment. The revision from $650B to $805B in a single reporting cycle is that loop, visible.

One figure makes the forward case more clearly than any projection. xAI has been running at roughly 11% GPU utilization on a fleet of 550,000 Nvidia cards; Meta and Google sit at 43–46%. The gap is not waste — it is pre-positioning. The organizations acquiring compute faster than their current workloads can saturate are betting the efficiency gains that drive consumption are still ahead. That also means the 11% figure reveals something the $805B number alone doesn’t: the acceleration documented so far is pre-Jevons. What happens when the efficiency curve hits the deployment base already built is the next chapter.

The Reckoning Reaches the Central Banks

Sintra, June 2026 — The Highest Institutional Tier Catches Up

In April, the structural case rested on analyst commentary and financial-research analysis. By June 2026, it had reached the tier where regulatory action originates. At the European Central Bank’s annual symposium in Sintra, Portugal, one question dominated the gathering of the world’s top central bankers and economists: whether AI is a boon or a threat to the global economy. Several joined the chorus warning of global fallout if the boom turns to bust. The Bank for International Settlements separately warned that the AI spending surge risks going into reverse and tipping some economies into recession.

The financial-stability case laid out there restates, at the highest institutional tier, the diagnosis in the sections above — and adds three amplifiers to the original depreciation thesis.

Leverage on Both Sides

This is the amplifier the original diagnosis implied but did not name. Depreciation asymmetry explains why the assets are worth less than booked. Leverage explains why that repricing goes systemic rather than staying contained — a downturn that feeds on itself through forced selling and margin calls rather than a quiet series of write-downs. The scale is now quantified: AI-related debt issuance is projected to more than double to nearly $570 billion in 2026, as hyperscaler capex outruns internal cash generation and pushes the build-out into the credit markets for the first time.

Concentration — No Firebreak

A repricing in a handful of names is a repricing of every index fund, pension, and investment-grade credit book. There is no diversification left to absorb the shock. This is the portfolio-level counterpart to the observation that AI now accounts for the majority of recent U.S. GDP growth: at that scale, a slowdown in AI capex is a slowdown in the economy itself.

Displacement-Driven Demand Collapse

Slok also laid out a two-scenario trap, both branches carrying downside. If AI succeeds wildly, it displaces jobs, consumer spending falls, and a recession follows. If adoption disappoints, the capex was built on returns that never arrive. Neither path requires the technology to fail. The disappointing-adoption branch is the macro expression of the readiness gap — returns are structurally delayed because organizations cannot absorb the capability at the pace the capex assumes.

The Counter-Weight in the Room

The consensus was not unanimous, and the diagnosis is stronger for saying so. Fed chairman Kevin Warsh refused the pessimistic framing — he put the revolution in its “first or second inning” and said he anticipates greater prosperity. Bank of Canada governor Tiff Macklem held the historically-literate middle: the internet proved more transformative than anyone imagined, and the dot-com bubble still happened. That is the exact distinction this paper depends on. The technology can deliver and the market can misprice the asset and the financing structure around it. Both can be true at once.

A Diagnosis, Not a Prescription

The Capex Reckoning is not a prediction about what might happen to AI infrastructure investment. It is a description of what is already happening — documented in financial analysis, revised upward by the markets within weeks of first publication, and now voiced at the level of the IMF, the BIS, and the European Central Bank.

The six forcing functions — economic obsolescence, supply foreclosure, operational drag, architectural uncertainty, accounting distortion, and geopolitical fracture — are not independent risks to be managed separately. They are a convergent structural shift. And the Sintra consensus has now extended that shift outward through two amplifiers: leverage, which makes the repricing systemic, and concentration, which removes the firebreak that would otherwise contain it.

The SaaS ARR reckoning happened when the market realized recurring revenue was not as safe as priced. The Capex Reckoning is happening now — as the market begins to realize that infrastructure investment is not as productive as booked.

This paper does not tell any enterprise where to place its infrastructure. That decision depends on factors specific to each organization’s data, workloads, regulatory posture, and competitive position — and the Sintra warnings about hyperscaler leverage and cloud concentration are a reminder that no placement is risk-free in either direction. What the paper does insist on is the diagnosis: the assumption underneath the balance sheets and the valuations is breaking.