AI Infrastructure Thesis Faces Capex Skepticism Even as Power Deals Accelerate

What changed

The past week has produced a sharp divergence in signals about AI infrastructure demand. On the supportive side, Microsoft signed a 20-year power deal with Chevron to secure electricity for AI data centers, marking a structural commitment to long-duration energy supply. SpaceX simultaneously signed a $6.3 billion compute deal with Reflection AI, with Reflection paying SpaceX $150 million per month starting July 1 for access to Nvidia GB300 chips inside the Colossus 2 data center. A third unnamed firm signed a 15-year, $2.6 billion AI power lease, according to reporting on the electricity-as-bottleneck thesis. Super Micro Computer introduced a new platform accelerating AI server backlog growth and gained 28% over two days on the announcement, notching its best two-day performance since May 2025.

Counterbalancing these developments, however, are material headwinds. Alphabet fell 6% and Amazon fell 4% on June 22 amid "AI capex anxiety across the hyperscalers," signaling investor concern that the pace and ROI of data center spending may be unsustainable. An AI pricing shock affecting OpenAI, Anthropic, and Microsoft emerged on June 18, suggesting margin compression in the model-as-a-service layer. Microsoft faces a June 30 earnings call that markets are treating as a potential inflection point for capex guidance. Separately, a shareholder lawsuit against Microsoft over cloud business expenses and AI spending was filed in Seattle federal court, reflecting accumulated investor skepticism about capex discipline.

Why it matters

The power-deal announcements materially strengthen the thesis's core mechanism: the bottleneck has genuinely shifted from semiconductor availability to electricity supply. Microsoft's 20-year Chevron commitment is not a short-term procurement contract—it is a decades-long infrastructure bet that locks in power supply at a time when grid capacity is the binding constraint on AI data center expansion. The $6.3 billion SpaceX-Reflection deal and the $2.6 billion power lease both validate that hyperscalers are willing to pay sustained, multi-billion-dollar premiums to secure reliable electricity. This is thesis-supporting evidence because it proves demand for AI compute is durable enough to justify long-duration capital allocation, not a temporary spike.

However, the simultaneous capex anxiety among investors introduces a critical counterweight. When Alphabet and Amazon both fell 4–6% on the same day over capex concerns, the market was signaling that the sustainability of hyperscaler spending is now in question, not the near-term demand for chips or power. This is a shift in the nature of the risk. The thesis assumes that hyperscalers will continue to deploy capital at scale; if investors believe capex cycles are peaking or that ROI on AI infrastructure is deteriorating, the entire buildout narrative contracts. The AI pricing shock reported on June 18 compounds this concern: if model providers are cutting prices, the revenue side of the ROI equation weakens, making it harder to justify continued infrastructure investment.

The Super Micro Computer rally on new platform announcements is a positive signal for the hardware layer of the thesis, but it must be weighed against the broader semiconductor selloff that has dominated the file since early June. SMCI's two-day 28% gain is a tactical bounce, not a reversal of the sector's structural repricing.

Opposing sources and risks

The capex anxiety narrative is the most material opposing signal. Multiple sources in the file describe investor concern that hyperscaler AI spending is unsustainable: the June 22 selloff in Alphabet and Amazon explicitly cited capex concerns, Oracle fell post-earnings on June 11 after AI spending guidance sparked cash flow concerns, and a June 12 article warned that backlash against data centers could show up in hyperscaler earnings reports. The AI pricing shock on June 18 is a direct threat to the ROI calculation that justifies capex in the first place. If frontier AI models are commoditizing and pricing is falling, hyperscalers have less revenue to offset the enormous capital outlays required for data center buildout.

A second risk is regulatory and grid-capacity constraints. A June 10 expert warning noted that America's grid is so far behind that blackouts are coming even without AI, suggesting that the electricity supply bottleneck may not be solvable through private power deals alone. If grid infrastructure cannot scale fast enough, even Microsoft's 20-year Chevron deal may prove insufficient to meet demand.

Third, the shareholder lawsuit against Microsoft over cloud business expenses and AI spending, filed in June, reflects accumulated investor skepticism about capex discipline and cash flow impact. If this lawsuit gains traction or if other institutional investors begin to question the capex thesis, it could trigger a broader repricing of hyperscaler valuations independent of fundamental demand.

What to watch

Third, monitor grid-capacity announcements and regulatory approvals for new power infrastructure. The Microsoft-Chevron deal is a private solution, but if regulatory or physical constraints prevent grid expansion, the electricity bottleneck could become a hard ceiling on data center growth.

Fourth, track Super Micro Computer's execution on its new platform and backlog conversion. SMCI's recent rally is contingent on the company delivering on the accelerated AI server backlog promise; any guidance miss would undermine the hardware layer of the thesis.

Finally, watch for additional hyperscaler power deals or capex guidance. Each new long-duration power agreement strengthens the thesis; each capex reduction or ROI warning weakens it.

Related Arbora context

This thesis is closely related to concept-megacap-tech-ai-monetization, which tracks the divergence in AI monetization credibility across megacap tech. The capex anxiety now affecting Alphabet and Amazon directly reflects investor doubt about whether these firms can monetize their infrastructure investments at a rate that justifies the spending. If the megacap monetization thesis deteriorates, the infrastructure buildout thesis loses its demand anchor.

concept-custom-silicon-ai-cloud-challenger-chips is relevant because Broadcom and AMD's custom ASIC offerings represent an alternative to Nvidia's GPU monopoly. If custom silicon gains share, it could extend the infrastructure buildout cycle by lowering the cost of compute and making capex more efficient—a thesis-supporting dynamic. Conversely, if custom silicon fails to scale, Nvidia's pricing power remains intact, which could accelerate capex exhaustion.

concept-ai-model-export-controls-sovereign-ai-access-risk is a second-order risk. If geopolitical restrictions on AI model access tighten, hyperscalers' international revenue from AI-as-a-service contracts could face a structural ceiling, reducing the ROI case for global data center expansion.

Sources

This article is research notes, not financial advice.