ED ZITRON AND DEAN BAKER: Much of the AI hardware already bought is not yet powered or installed, so chip sales overstate working capacity and real demand

The Gist

Big tech companies have bought far more AI chips than they have plugged in, mostly because new data centers can't get power fast enough. So chip sales make AI demand look bigger than the capacity actually running. Chips that sit waiting also go out of date. This packet steelmans the arguments made by Ed Zitron and Dean Baker in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

A large share of the AI chips sold since 2024 appears not yet to be installed in powered data centers, so headline chip sales and capital spending overstate how much working AI capacity exists and how much paying demand it serves, and chips waiting for power lose economic value as newer generations arrive.

Premises

1. Reporting based on internal data indicated Microsoft had about 2.2 million AI chips in service, and Bloomberg later reported that only about 2 of Microsoft's roughly 12 gigawatts of data center capacity is centered on AI-specific chips.
2. Microsoft has spent about $265 billion on capital expenditures since 2022, while the GPUs reported as in service are worth on the order of $50 to $60 billion.
3. Construction-in-progress balances across major hyperscalers, neoclouds, and colocation firms exceed $370 billion, and a material share of that likely consists of purchased chips and servers awaiting buildings or power.
4. A Morgan Stanley estimate reported in the Financial Times found that more than half of the GPU servers expected to be sold in 2026 through 2028 may have no powered place to be installed.
5. Companies do not break out installed versus uninstalled AI hardware, so investors cannot see how much spending has become working capacity.
6. AI chips lose economic value quickly as newer generations arrive, so chips that wait a year or more for power may be worth much less when they finally run.

Assumptions

• The estimate that over $200 billion of chips are uninstalled is the speaker's own calculation from construction-in-progress balances and an assumed split between hardware and buildings; it is not a disclosed figure.
• Some lag between buying hardware and billing for it is normal; Amazon has said it lays out cash 6 to 24 months before it can bill customers.
• Measured GPU failure rates are low (well under 1 percent a year in one large study), so the concern is economic obsolescence, not chips physically burning out within two years.
• The Morgan Stanley estimate covered 2026 through 2028 and is a projection of a power shortfall, not a count of idle chips.

Analysis

Argument type: Inductive.

Potential Fallacies

• Apples-to-oranges comparison (residual treated as idle hardware) (Premises 1 and 2) — Cumulative capex ($265B) includes buildings, land, power and cooling equipment, networking, CPUs, storage, leases, and non-AI cloud growth. Subtracting the value of in-service GPUs therefore does not isolate uninstalled chips. Likewise, 2 GW of 12 GW describes the mix of capacity types, not how many purchased chips lack power. Without a stated breakdown, the gap is consistent with idle chips but equally consistent with ordinary non-chip spending.
• Projection treated as present fact (Premise 4 and the conclusion) — The Morgan Stanley estimate is a forward-looking forecast of a 2026-2028 power shortfall (as the assumptions acknowledge). The conclusion makes a present-tense claim about chips already sold since 2024. A forecast shifts the odds of future stranding more than it demonstrates current idleness, and buyers may adjust orders or delivery timing in response.
• Proxy conflation (CIP as a stand-in for idle chips) (Premise 3) — Construction-in-progress mostly holds buildings, electrical and cooling systems, and equipment staged for imminent commissioning. The 'material share' that is chips is an assumed split, and the pooled figure mixes hyperscalers, neoclouds, and colocation firms with very different asset compositions. This makes the headline dollar figure highly uncertain.
• Inference from absence of disclosure (Premise 5) — That companies do not report installed versus uninstalled hardware shows the quantity is unknown to outsiders. It does not show the quantity is large. Opacity is equally expected whether the idle share is small or large, so it justifies caution and calls for disclosure, not a particular estimate.
• Unsupported leap from supply-side lag to demand weakness (Conclusion (paying demand clause)) — The premises concern power, buildings, and installation timing. None measures customer revenue, utilization, backlog, or rental prices. Uninstalled chips can coexist with, or even signal, demand that exceeds powered capacity, so the 'paying demand' clause of the conclusion does not follow from the evidence.
• Single-firm and shifting reference-class generalization (Premises 1, 2, 4 and the conclusion) — Microsoft's figures, which cover capex since 2022, are used to support an industry-wide claim about chips sold since 2024, alongside a 2026-2028 projection. Firms differ in their stage of buildout, chip mix (including custom silicon), and accounting, so the premises do not map cleanly onto the conclusion's population and period.
• Missing baseline (base-rate neglect) (Premises 2 and 3) — The assumptions concede that lag between purchase and billing is normal (6-24 months at Amazon), but the argument does not say how much of the observed gap exceeds that norm. Without a baseline, large spend-to-service gaps cannot be distinguished from a healthy pipeline in a fast-growing buildout.

Counterarguments

These challenges are not ranked.

• Conclusion (paying demand clause) (High impact) — If power and buildings are the binding constraint, uninstalled chips indicate demand exceeding powered capacity, not overstated demand. Evidence such as cloud backlogs, remaining performance obligations, capacity-constrained commentary, and rising rental prices would point the opposite way from the conclusion. No premise offers demand-side data.
• Premise 2 (High impact) — Capex is mostly non-GPU infrastructure and includes pre-AI cloud expansion, leases, and depreciated assets. The gap between $265B and $50-60B is the expected signature of a build-out, not evidence of idle chips.
• Premise 3 (High impact) — CIP grows in any capex boom and is dominated by shells, electrical gear, and cooling. Colocation CIP is mostly buildings, and equipment staged for commissioning is classified there while days from service. The chip share could be far below the speaker's assumed split, so the $200B+ figure may be much too high.
• Conclusion (High impact) — Hyperscalers rationally pre-buy scarce GPUs to secure allocation and sequence construction over multi-year horizons. A lag measured in months, consistent with the normal 6-24 month range, is a modest timing cost relative to the value of securing supply, and does not mean headline figures mislead.
• Premise 6 (Medium impact) — Previous-generation GPUs retain rental and resale value, and are used for inference and lower-tier workloads, especially under supply constraints. Value loss from waiting may be modest and depends on pricing, not merely on newer chips existing.
• Premise 4 (Medium impact) — A projected shortfall can be resolved through delayed deliveries, reduced orders, on-site generation, grid investment, or redeployment to powered sites. Projections also depend on assumptions about shipments that respond to the shortfall itself, so strandedness is not a foregone outcome.
• Premise 1 (Medium impact) — The 2.2M chip count may be stale, partial, or narrowly defined, and Microsoft may use leased or partner capacity not captured in its own figures. A low AI share of total gigawatts is the expected state of a fleet in which AI is the newest and fastest-growing layer.

Suggested Improvements

• Separate the three claims — Split the argument into (1) a backlog of uninstalled chips exists, (2) waiting erodes its value, and (3) headline figures overstate paying demand, with distinct evidence for each. Drop or soften the demand claim unless demand-side data is provided. The demand conclusion does not follow from supply-side evidence, and bundling it weakens the better-supported capacity claim.
• Establish a baseline for normal lag — Compare current CIP-to-in-service conversion times and chips-per-MW energization ratios against historical cloud build-outs and against Amazon's 6-24 month range, showing whether the delay is lengthening. Without a benchmark, a healthy pipeline cannot be distinguished from a stuck one.
• Decompose capex and CIP — Use 10-K/10-Q footnotes to split CIP and capex by asset class, and present the >$200B estimate as a range with a sensitivity analysis over plausible hardware/building splits. The central dollar figure rests on an undisclosed assumption. A transparent range would reflect its true uncertainty and make it checkable.
• Reconcile units and time frames — Match vendor shipment data (units sold since 2024) against energized AI capacity (MW times chips per MW) for the same period and the same firms, instead of mixing chips, gigawatts, and cumulative dollars. A like-for-like reconciliation would test the claim directly and remove the apples-to-oranges objection.
• Add demand-side evidence — Cite utilization, contracted backlog, revenue per installed GPU, and rental price trends, and address why power-constrained supply would indicate overstated, rather than unmet, demand. These are the data that would actually bear on the 'paying demand' claim.
• Quantify obsolescence — Use secondary-market and rental price data by GPU generation to estimate value loss over a realistic wait, and engage the depreciation-life debate (5-6 years versus 2-3 years). This would turn a plausible assertion into a measured magnitude and address the counterevidence that older GPUs still earn income.
• Broaden beyond Microsoft — Add company-by-company evidence for other hyperscalers and neoclouds, noting differences in custom silicon and disclosure. A single-firm snapshot cannot support an industry-wide claim.

Scenario Tests

• Company filings or supply-chain data show that most shipped GPUs are energized within about 3-9 months of delivery. (Challenges) — The backlog would be a normal pipeline, the obsolescence loss would be small, and the claim that headline sales overstate working capacity would be right only in a trivial timing sense.
• Reconciling GPU shipments against energized AI capacity shows a cumulative shortfall that grows over several quarters. (Supports) — A widening gap would indicate a binding power or construction constraint, strengthening the capacity-overstatement claim and the case for disclosure.
• Cloud providers report sold-out capacity, rising backlogs, and rising rental prices for older GPUs while chips wait for power. (Challenges) — Uninstalled chips would reflect excess demand constrained by supply, undermining the demand-overstatement claim and the rapid-obsolescence premise, although the capacity-lag claim could survive.
• Hyperscalers take large impairments or accelerate depreciation on GPU fleets, or leveraged neoclouds struggle to service debt on chips that were slow to earn revenue. (Supports) — This would show that obsolescence during waiting is economically material and that the opacity was masking a real problem.
• The Morgan Stanley power-shortfall projection fails to materialize because of on-site generation, grid investment, or order deferrals. (Challenges) — The forward-looking support for stranded chips would weaken, and the market's adaptive response would be shown to limit the problem.
• Companies begin disclosing installed versus uninstalled accelerators, and the uninstalled share is large and persistent. (Supports) — This would directly validate the core claim and show that the speaker's estimate was directionally right, though the dollar magnitude could still differ.

Coherence & Relevance

The argument is a cumulative-case inductive argument whose lines of evidence (Microsoft's in-service ratio, aggregate CIP, an analyst projection, disclosure opacity, and obsolescence) converge on a plausible, but unquantified, installation backlog. The premises are individually indirect, partly overlapping, and cover mismatched populations and time frames, so they support the existence of some backlog much better than a 'large share' or a specific dollar figure. The stated assumptions candidly limit the claim and are treated as given. The conclusion's final reach, that paying demand is overstated, is not connected to any premise and could be contradicted by the same facts if power is the binding constraint. A tighter version would claim a meaningful, unquantified backlog that erodes in value and is not visible to investors, and would call for better disclosure.

• Reporting based on internal data indicated Microsoft had about 2.2 million AI chips in service, and Bloomberg later reported that only about 2 of Microsoft's roughly 12 gigawatts of data center capacity is centered on AI-specific chips. (Weak) — Provides a count and a capacity mix, but no denominator of total chips purchased. The figures are of different types and dates, and they do not show that purchased chips are uninstalled.
• Microsoft has spent about $265 billion on capital expenditures since 2022, while the GPUs reported as in service are worth on the order of $50 to $60 billion. (Moderate) — Directionally suggestive of a gap between spending and working GPUs, but the residual includes many non-chip categories. No decomposition is given, and the time frame (since 2022) differs from the conclusion's 'since 2024'.
• Construction-in-progress balances across major hyperscalers, neoclouds, and colocation firms exceed $370 billion, and a material share of that likely consists of purchased chips and servers awaiting buildings or power. (Moderate) — Relevant to the existence of a backlog, but the chip share is an assumed split, 'material' is unquantified, and the pooled firms have heterogeneous asset mixes.
• A Morgan Stanley estimate reported in the Financial Times found that more than half of the GPU servers expected to be sold in 2026 through 2028 may have no powered place to be installed. (Moderate) — Strong on power as a binding constraint, but it concerns future sales (2026-2028) and not chips already sold since 2024, and it is a projection, not an inventory count.
• Companies do not break out installed versus uninstalled AI hardware, so investors cannot see how much spending has become working capacity. (Moderate) — Explains why the central quantity must be estimated and supports the case for disclosure, but offers no evidence about the size of the idle stock.
• AI chips lose economic value quickly as newer generations arrive, so chips that wait a year or more for power may be worth much less when they finally run. (Moderate) — Supports the final clause about value loss, conditional on long waits. It is independent of how many chips are idle and has no data on wait durations or resale and rental prices. It does not connect to the demand claim.

View this argument on LogicFirst.ai