Hyperscaler capex near $780 billion is consuming a growing share of cash flow, a16z charts show

On that chart's figures, spending grew about 73% from 2024 to 2025 and is on course to grow about 90% from 2025 to 2026 ^[1][2]. The report's other chart rounds this year to $780 billion and says annual spending could cross $1 trillion soon after ^[2]. The two charts are $10 billion apart for the same year ^[3].

Capex comes out of operating cash flow. Free cash flow is what remains after it ^[20]. For most of the boom, the hyperscalers' existing businesses generated enough cash to cover the build ^[16]. Epoch AI's growth rates put a date on the end of that. If they held, capex would overtake operating cash flow around the third quarter of 2026 ^[17]. Past that point free cash flow goes negative, and the shortfall has to be funded from reserves or from outside the business ^[4].

a16z's explanation for the A100 rental floor is that cheaper compute draws more use, so older chips stay busy while total demand grows ^[18]. Runtimewire points out that the value of installed hardware depends partly on how much demand expands before the next generation arrives ^[19]. For the rental figure to apply to a specific team, the rates a16z tracked would have to match the contract term and region that team buys. Runtimewire also notes the report does not establish that every chip, operator or data center keeps its value, and a16z says the story is still unfolding ^[8].

The coverage backs the physical-capital reading on where the money is going. According to a16z, capital is moving into semiconductors, power, networking, construction, debt markets and skilled labor ^[9]. For a product sold as the cloud, the bill looks a lot like a construction budget. Neither write-up reports power prices, grid connection times or wage figures from the report. The case that electricity and skilled labor will cap what teams can build is an inference from capital flows. The only price series in the coverage that touches a team's budget is GPU rental ^[7].

The firm's own disclosures say some of a16z's data comes from third parties and portfolio companies and has not been independently verified ^[11]. Runtimewire notes that the report's figure for tech's share of S&P 500 earnings growth this year through late August, about 76%, comes without the calculation or dataset behind it ^[14]. The author, David George, leads the firm's Growth practice ^[13]. His thesis matches the physical buildout the firm's new fund is set up to back ^[12].

In my view, flat prices for older GPUs are the right planning assumption for a team that rents, until a rental series turns down. A team that owns A100s sits on the other side of the same price.