Microsoft, Meta and Google climb to $725 billion in AI capex: the datacenter race enters record territory

Microsoft, Meta and Google climb to $725 billion in AI capex: the datacenter race enters record territory

🔎 Four companies, $725 billion, one single question

The figure is so large it becomes abstract: roughly $725 billion in combined capital expenditure (capex) for 2026 across Amazon, Microsoft, Google and Meta, according to data reported by Reuters and picked up by tech-insider.org. These same companies had invested $388 billion in 2025 — already a record. They are doubling down within twelve months.

On the menu: around $200B at Amazon, ~$190B at Microsoft (over its fiscal year), $180–190B at Google, $125–145B at Meta. Behind the headline numbers lies a physical reality: gigawatts, land, transformers, servers — and a collective bet on demand for the entire decade.

Why now? Because demand, at last, can be measured in contractual billions: $627B in commercial RPO at Microsoft, $37B in AI run-rate, a cloud backlog of more than $240B at Google. And because the skeptics — Michael Burry and his "too big to fail" oligopoly, Ed Zitron and his anti-bubble crusade — have never had so much to grab onto.

So the central question is simple: does demand justify the spending? Short answer: largely yes for 2026. The risk lies elsewhere — in the concentration of demand, the cost of financing, and a bottleneck nobody talks about enough: electricity.

The Essentials

• ~$725B in combined capex in 2026 for Amazon, Microsoft, Google and Meta (Reuters), versus $388B in 2025: a near-doubling.
• Breakdown: Amazon ~$200B, Microsoft ~$190B (fiscal year), Google $180-190B, Meta $125-145B.
• Demand is contracted: $627B in RPO at Microsoft, $37B in AI run-rate, Alphabet's cloud backlog above $240B (+55% in a single quarter).
• The bottleneck is electrical: $80B of Azure backlog is waiting for megawatts, not customers.
• The skeptics (Burry, Zitron) point to a "too big to fail" oligopoly and Meta's free cash flow projected to fall by ~90% (Barclays).
• Our verdict: the spending is justified for 2026; the risk plays out in 2027, around demand concentration and asset lifespans.

Recommended Tools

You don't have $725 billion — and that's just fine. The paradox of this race is that it makes AI cheaper for everyone. Here's how to take advantage of it without committing a single euro of capex.

To go further, our guide to free AI APIs details the quotas and limits of each offering.

The numbers: $725B… but which one should you count?

The Big Four of cloud will spend roughly $725B in 2026. But depending on the source, you'll also come across $630B, $690B, or $700B — and all of these estimates hold up. They simply aren't measuring the same thing.

Sources: tech-insider.org, Reuters, Futurum, CNBC, Data Center Richness. Since Amazon is not broken out separately by Data Center Richness, its 2025 budget is inferred from the $388B total.

Why such discrepancies? Three reasons. First, the scope: Data Center Richness puts planned capex for the four hyperscalers at $630B at most (+62% vs 2025), Futurum includes Oracle and its ~$50B for a total of $660-690B across five players, and CNBC cites "nearly $700B" for the four. Reuters settles on $725B.

Next, the fiscal year. Microsoft closes its year in June: its ~$190B covers an offset fiscal year, whereas Data Center Richness's $110-120B is on a calendar basis. Comparing the two without care is like comparing a kilogram to a pound.

Finally, finance leases and guidance ranges: a data center can be accounted for as a purchase or a lease, and Meta communicates in ranges — $115-135B per Futurum, up to $145B per the highest estimates.

Let's add one detail that makes your head spin: Microsoft spent $37.5B in the last quarter alone (Futurum). The giants' quarterly pace now exceeds the digital infrastructure budget of most countries.

My advice: when faced with a headline about AI capex, check three things — how many players are counted, over which fiscal year, and with which bounds. Without that, the debate is fruitless.

Does demand justify the spend? The optimists' case

Yes — and unlike 2023, demand is measured in contractual dollars, not slide projections.

The best indicator is called RPO, for remaining performance obligations: revenue from signed contracts not yet recognized. Microsoft reports $627B in commercial RPO, to which you can add a $37B AI run-rate — the AI business alone is already running at the level of a major CAC 40 company's annual revenue.

Google is on the same trajectory: its cloud backlog jumped 55% in a single quarter, beyond $240B (Futurum). And the most telling detail comes from Azure: of the $80B in unfulfilled orders, the problem isn't demand but power availability. When your customers are waiting for megawatts rather than discounts, you're no longer in a demand bubble.

On the compute-buyer side, the revenue exists too: Anthropic crossed $47B in revenue run-rate and overtook OpenAI. Enough to back massive cloud contracts with real monetization, not just venture capital.

The technology itself, finally, plays against the savings: reasoning models and agents consume orders of magnitude more compute per query than simple chat. GPT-5.5 (98.2 on the agentic index), Gemini 3 Pro Deep Think (95.4), or Claude Opus 4.7 (94.3) "think" before answering — and every step of thinking is billable GPU time. An agent working through the night consumes more tokens than a human could ever write.

And the loop closes: these AI revenues fund capex, which increases capacity, which attracts new customers. This is exactly the virtuous mechanism the skeptics contest — we'll get to that.

My take: RPO isn't revenue, and an order book can age badly. But it's the best leading indicator available — and it leaves little room for the invented-demand hypothesis.

The skeptics' camp: "too big to fail," depreciation, and falling cash flow

The critics don't deny the 2026 demand. They worry about the mechanics — and about what happens the day it seizes up.

Michael Burry sums up the thesis in one phrase: an oligopoly that is "too big to fail." Four balance sheets now carry a massive share of stock market growth. At this stage, halting capex would become a macroeconomic event — and none of the four has any interest in being the first to deviate. It's game theory applied to datacenters: everyone can see the endgame is risky, and everyone speeds up anyway.

Ed Zitron attacks the economic foundation: nothing so far proves that AI revenues will cover the cost of the capital committed. His strongest argument concerns depreciation — a cutting-edge GPU can be outdated within a few years, while a datacenter depreciates over decades. If technology cycles accelerate, the depreciation schedules will turn out to be optimistic.

The most uncomfortable figure comes from CNBC: Barclays projects a drop of nearly 90% in Meta's free cash flow, with capex potentially reaching $135B. Meta monetizes AI through advertising — a bridge between generative models and ad revenue that remains to be proven at this scale.

The markets' counter-argument: the four leaders held more than $420B in cash at the end of last quarter. They can fund the race without taking on massive debt. True — but a cash cushion doesn't answer the return-on-investment question; it merely postpones the reckoning.

My take: Burry and Zitron are right about the mechanics; they may be wrong about the outcome. An oligopoly investing massively in the face of real, contracted demand is not 2000 and its sunken fiber optics — it's a bet on who captures the demand of 2030. The risk of capital destruction exists; a bloc-wide collapse would require demand to disappear, and nothing indicates that.

The real bottleneck: electricity, not chips

The scarce resource in this race is neither the GPU nor capital: it's the megawatt.

Futurum puts it bluntly: Microsoft's $80B unfulfilled Azure backlog is "a matter of electrical availability, not demand." Translation: Microsoft has more customers than outlets.

This detail changes the entire reading of capex. A datacenter connects to the grid on timelines measured in years, and electrical capacity is reserved far in advance. When a slot opens up, you have to commit — hence capex announcements that look like bets, but are mostly megawatt reservations.

This also explains why budgets are doubling even as GPU shortages ease: the constraint has shifted downstream, to grids, transformers, and energy contracts. Hyperscalers aren't spending $725B because they can — they're doing it because every megawatt secured today is contracted revenue tomorrow.

For enterprise customers, it's also a signal: time-to-service becomes a cloud selection criterion, on par with price. A lingering backlog means deferred revenue — and an open door for regional competitors.

The lesson is brutal for cloud challengers: the game is now played on access to energy. And that's exactly where China, which plans its grid at the national level, becomes a structural rather than commercial threat.

What About China? ByteDance's $70B and the Open-Weight Offensive

Faced with America's $725B, China cannot match the same balance sheets. It compensates with massive capex on the platform side and efficiency on the model side.

ByteDance is thus preparing $70 billion in AI capex for 2026 (our analysis). That's ten times less than the Big Four, but a colossal budget for a single company — and it fits within a national strategy where the power grid is planned at the state level.

The other front is that of open models. Moonshot AI raised $2B while Kimi K2.6 dominates open-weight and China accelerates: in self-hosting, Kimi K2.6 scores 88.1 on the agentic index, ahead of GPT-5.4 (87.6), with Z.AI's GLM-5 (Reasoning) at 82. Models trained and run on infrastructure constrained by chip export controls.

For the American capex thesis, this is the most uncomfortable point: if algorithmic efficiency catches up with computing power, the hyperscalers' moat shrinks. An architecture that cut inference costs by an order of magnitude would render part of these datacenters superfluous. It's the scenario skeptics hint at implicitly — and one no one knows how to date.

My take: the Chinese risk isn't winning the datacenter race. It's making it pointless.

What This Means for You: Rented AI Has Never Been This Cheap

The paradox of this race: the more the hyperscalers spend, the cheaper AI gets for those who consume it.

$725B in capex means inference capacity installed far beyond current demand — hence permanent downward pressure on prices. Every player has to fill their datacenters; developers reap the benefits: more generous free tiers, falling per-token prices, open models that run anywhere.

Concretely, you can build a complete AI product without spending a cent on infrastructure: a brain from Groq, Google, or OpenRouter — see our selection of free AI APIs, a hosted interface for a few euros a month at Hostinger. The giants' overinvestment subsidizes your MVP.

Three habits for 2026: start with the free tiers before any commitment, avoid annual contracts as long as inference prices keep falling, and keep your applications portable from one model to another. The window of opportunity is during overcapacity, not after.

❌ Common Mistakes

Mistake 1: comparing aggregates without checking the scope

$630B, $690B, $700B, $725B: all these figures are circulating, and they don't describe the same reality (4 or 5 players, fiscal or calendar year, leases included or not). Before drawing any conclusion, check who is being counted, over what period, and with what guidance boundaries.

Mistake 2: confusing capex with "burned" cash

A data center is a depreciable asset, not an operating expense. The real debate is about the useful life of equipment in the face of GPU refresh cycles. Judge these companies on their RPO and their revenues, never on capex alone.

Mistake 3: treating 2026 guidance as set in stone

The ranges shift every quarter — Meta went from $72B spent in 2025 to three-figure guidance, revised with each announcement. Building an investment thesis or a business plan on a single figure is a methodological error.

Mistake 4: ignoring demand concentration

$627B in RPO, yes — but signed with a small number of AI customers. If one of them slows down, the backlog ages without disappearing. The metric to watch in 2027 is not total capex: it's the diversification of compute buyers.

❓ Frequently Asked Questions

Why do we talk about $725B when other sources cite $630B?

Because the scopes differ: Reuters aggregates the four hyperscalers (~$725B), Futurum counts five players including Oracle ($660–690B), and Data Center Richness keeps the planned capex of the Big Four alone (up to $630B, +62% vs 2025). Microsoft's offset fiscal year and the treatment of leases explain the rest.

Does demand really justify this spending?

For 2026, largely yes: $627B in commercial RPO at Microsoft, a $37B AI run-rate, a cloud backlog above $240B at Google, and a $47B revenue run-rate at Anthropic. The doubt concerns 2027 and beyond: demand is real, but young and concentrated among a few buyers.

What does Microsoft's $627B RPO mean?

Remaining Performance Obligations: revenue from contracts already signed but not yet recognized. It's contracted demand, not a forecast. Its limits: the execution timeline (electricity delays Azure deliveries) and the financial health of the signatories. It is nevertheless the market's best leading indicator.

Is Michael Burry right about the "too big to fail" risk?

On the mechanics, yes: four balance sheets support the market, and an abrupt halt in capex would be a macroeconomic shock. But these companies held more than $420B in cash at the end of last quarter, along with diversified business models. Systemic risk exists; its materialization would require a collapse in demand, which has not been observed to date.

How can you take advantage of this race without a datacenter budget?

By renting everything: the free tiers of Groq, Google, and OpenRouter cover prototyping, a few euros a month in hosting carries your applications, and open models like Kimi K2.6 run self-hosted. The overcapacity created by the $725B translates directly into falling inference prices.

✅ Conclusion

At $725 billion a year, the datacenter race is no longer a bet on AI: it's a bet on the power grid and 2030 demand — and while the giants battle it out, your best move is still to leverage their spare capacity, starting with our guide to free AI APIs.