Google's AI Infrastructure Spend in 2026: $185B Capex, TPU v7, and the Cloud Bet

Alphabet is spending $180-190 billion on capital expenditure in 2026 — more than double the $91.4 billion it spent in 2025. That's the short answer. The longer answer is that Google just raised $80 billion in equity to help pay for it, and Google Cloud's backlog nearly doubled in a single quarter to $460 billion, which is the number management is using to justify all of it.

I've watched a lot of "we're doubling down" earnings calls over the years, but Alphabet raising external capital to fund capex — after decades of self-funding everything from cash flow — is a real signal. This is a look at where the $185B midpoint is actually going, what TPU v7 Ironwood changes, and whether the demand numbers back up the spend.

How Much Is Google Spending on AI Infrastructure in 2026?

Alphabet's 2026 capital expenditure guidance sits at $180-190 billion, raised during the Q1 2026 earnings call from an initial $175-185 billion range given back in February. That is more than double the $91.4 billion Alphabet actually spent in 2025, which itself was already a sharp jump from prior years. Management has also flagged that 2027 capex is expected to increase further compared to 2026 — this is not a peak-and-plateau story, it's an accelerating one.

The stated use of funds is investing in AI compute capacity for Google DeepMind, meeting what the company calls "significant cloud customer demand," and strategic investments in other bets — plus improving the user experience and advertiser ROI across core Google services. In plain terms: TPUs, data centers, networking, and power, in that order of dollar volume.

Google's Capex Growth: 2024-2026

The trajectory is the story here more than any single-year number. Alphabet went from roughly $32B in 2023 capex to $52.5B in 2024, $91.4B in 2025, and now a guided $180-190B in 2026 — a near-doubling in each of the last two years running.

Figures are 2026 estimates blended from Alphabet 8-K earnings filings (SEC EDGAR), CNBC, and TechI reporting on capex guidance revisions through Q1 2026.

The $80B Equity Raise: Why Google Stopped Self-Funding

For most of the last decade, Alphabet funded its entire capex program out of operating cash flow — it never needed to raise external capital for infrastructure. That changed in 2026: alongside the higher capex guidance, Alphabet announced equity offerings totaling roughly $80 billion specifically earmarked to fund AI compute investment and meet what it calls "unprecedented customer demand."

That's a meaningful tell. A company generating well over $100 billion a year in free cash flow choosing to raise equity anyway means the capex number has outgrown what even Google's cash engine can comfortably absorb in a single year without straining the balance sheet or crowding out buybacks. It puts Alphabet in the same camp as Meta and Microsoft, both of which have also turned to debt or off-balance-sheet financing structures to fund 2026 AI infrastructure at this scale.

TPU v7 Ironwood: What the Chip Actually Does

A large share of the 2026 capex is going into Google's seventh-generation Tensor Processing Unit, Ironwood, which reached general availability on April 22, 2026. Each chip carries 192GB of HBM3E memory at up to 7.37 TB/s bandwidth and delivers 4,614 FP8 TFLOPS. A full pod scales to 9,216 chips for 42.5 FP8 ExaFLOPS — roughly 10x the peak performance of the prior TPU v5p generation and more than 4x the per-chip performance of TPU v6e (Trillium).

Google has committed to roughly 1 million TPU chips for 2026: an initial 400,000-unit phase worth roughly $10 billion in finished racks purchased through Broadcom, plus a further ~600,000-unit commitment tracked as roughly $42 billion in remaining performance obligations, rented out through Google Cloud. Reports suggest Google's internal forecast for the next TPU cycle has been revised up from about 2 million units to roughly 4 million — a step-up that's also driving a jump in optical module demand from an estimated 3 million modules in 2025 to around 20 million in 2026.

For how TPUs stack up against Nvidia and AMD silicon on cost and performance, see our AI hardware wars breakdown.

Is Google Cloud Growing Fast Enough to Justify the Spend?

The demand-side case rests on Google Cloud's numbers. Cloud revenue hit $20.0 billion in Q1 2026, up 63% year-over-year, with backlog (remaining performance obligations) nearly doubling quarter-over-quarter to over $460 billion. Gemini Enterprise paid monthly active users grew 40% quarter-over-quarter, and Gemini models are now processing more than 16 billion tokens per minute via direct API use, up 60% from the prior quarter.

CEO Sundar Pichai was explicit on the earnings call that Q1 revenue "would have been higher" if Google could meet all the demand it's seeing — meaning growth was capacity-constrained, not demand-constrained. That's the strongest argument management has for why capex needs to keep rising rather than plateau: every quarter of underbuilt capacity is revenue Google is visibly leaving on the table, not a hypothetical.

Google vs Microsoft, Amazon, and Meta: 2026 AI Capex Compared

Google's $185B midpoint is large but not the largest among the hyperscalers. Amazon is projected to spend around $200 billion in 2026, the biggest single number, largely for AWS data centers and Trainium chips. Microsoft is tracking toward roughly $190 billion, mostly for AI data centers and GPU/Azure compute tied to its OpenAI relationship. Meta guided to $115-135 billion, nearly double its $72 billion spent in 2025. Combined, the five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — have committed to $660-690 billion in 2026 capex.

For the full four-way spending race, see our Big Tech AI capex comparison, and for how our AI Valuations dashboard tracks the resulting price tags on AI-native startups riding this infrastructure wave.

What This Means for AI Startups and Investors

A $185B infrastructure build from a single hyperscaler reshapes the economics for every AI startup building on top of it, not just Google's own products. TPU v7 access through Google Cloud is now a real cost lever for model training and inference — Ironwood's 4x per-chip performance gain over TPU v6e means startups renting TPU capacity can plausibly cut training cost per model by a meaningful margin versus a year ago, if they can get allocation at all given the capacity constraints Pichai flagged.

For venture investors, the read-through is twofold. First, the $460B Cloud backlog is itself a leading indicator of enterprise AI adoption — companies don't sign multi-year compute commitments of that size speculatively, so it's a reasonable proxy for how much real production AI workload is coming, not just pilots. Second, the sheer scale of the capex race (Google, Microsoft, Amazon, and Meta combined committing $660-690B in 2026 alone) means compute is increasingly a scale advantage that only a handful of companies can underwrite, which pushes more of the AI value chain toward companies that rent rather than own infrastructure — exactly the API and platform layer our AI Valuations dashboard tracks.

There's a real risk on the other side too. If Cloud revenue growth decelerates from 63% while capex keeps compounding toward 2027, Alphabet's return on invested capital for this cycle craters, and the $80B equity raise starts to look premature rather than prescient — a scenario every LP underwriting AI-infrastructure-adjacent funds should be modeling, not just assuming away. For a broader view of how the capex cycle is playing out across power and physical build-out, see our AI data center power demand breakdown.

The Bottom Line

Google's 2026 AI infrastructure spend has genuinely outgrown its own cash flow for the first time in the company's history, which is why it went to equity markets for $80 billion rather than fund the entire $185B guide internally. The offsetting fact is that Google Cloud's $460B backlog and capacity-constrained growth are real, verifiable demand signals, not just capex justified by FOMO.

The number to watch through the rest of 2026 isn't the capex guide itself — it's whether Cloud revenue growth keeps accelerating in lockstep, or whether the backlog conversion slows once TPU v7 capacity actually comes fully online. If it does, Google looks smart for raising capital ahead of the curve. If Cloud growth decelerates while capex keeps climbing toward 2027, this is the same overbuild risk every hyperscaler is now underwriting at once.