Big Tech Is Waging a $650 Billion Capex War, the Founder's Weapon Is Cheap Inference

What Happened

For the first time, 2026 AI infrastructure spending has jumped to a different curve. The four hyperscalers, Amazon, Microsoft, Google, and Meta, have collectively earmarked about $650 billion in capex this year. Amazon guides to roughly $200 billion, Alphabet to $175-185 billion, Meta to $125-145 billion, and Microsoft to an annualized run rate near $145 billion. Meta spent $72.2 billion on capex in 2025, so it is more than doubling in a single year, and Zuckerberg has committed to at least $600 billion in U.S. AI infrastructure by 2028. On top of that sits Stargate: the joint venture of SoftBank, OpenAI, Oracle, and MGX plans to invest $500 billion in AI infrastructure over four years, with Masayoshi Son as chairman. SoftBank closed an additional $22.5 billion investment in OpenAI on December 26, lifting its stake to roughly 11 percent, and has pledged EUR 75 billion (about $87 billion) of data centers in France. Son told CNBC this cycle is 50 times bigger than the dot-com boom. Markets flinched: analysts warn Big Tech free cash flow could fall by as much as 90 percent in 2026 as capex outruns revenue.

What This Means for Founders

The square where you compete on capital is not yours to play. A $600 billion data center pledge leaves no room for a small team. But this capex has an opposite-facing effect. The compute that giant balance sheets are erecting is a deflation machine for the one input founders actually pay for: inference. GPT-4-class performance that cost $30 per million tokens in early 2024 now runs $2-3. That is a tenfold drop in two years. Open-weight models are another 10x to 100x cheaper on top of that, with DeepSeek-class APIs down to $0.14 per million input tokens through providers like Together, Fireworks, Groq, and DeepInfra, which now set the price floor that OpenAI, Anthropic, and Google compete against. On a cost curve that Big Tech is subsidizing by torching its own free cash flow, the founder’s edge moves from owning compute to capital efficiency. Don’t buy GPUs, rent them. Put a thin app on cheap tokens and dig your moat in distribution, workflow, and proprietary data, the same lean playbook YC has preached for a decade.

One thing needs a filter. Even with a 10x price drop, if token consumption rises 100x the bill grows anyway. Many startups built on LLMs still burn 40-60 percent of revenue on infrastructure. You don’t ride the cost curve for free; you architect on top of it, or the deflation never reaches your margin.

What You Can Do Now

Build for model portability. Split the work that genuinely needs a frontier model from the high-volume jobs open weights handle fine, namely classification, RAG, summarization, and bulk generation, and route the back rows to the cheapest supply. The 90 percent free-cash-flow warning is also a signal for founders: today’s cheap inference carries a subsidy that Big Tech is eating at a loss, and prices can snap back. Teams that bank user habit and a data moat while costs are at the floor are the ones that survive the rebound.