Business Insider published a report based on an internal audio recording of an AWS meeting from September 2026: CFO John Felton jokingly “congratulated” Amazon’s head of AI infrastructure, Prasad Kalyanaraman, on driving the company to negative free cash flow. Without irony, the situation looks like this: in 2026, Amazon is allocating $220 billion to build data centers for AI, and for the first time in a long time, the company is in the red despite record growth in its cloud division. AWS leadership confirms that this situation will persist for another couple of years.

image
image

What happened

In the recording of the internal AWS meeting from September 2026, which Business Insider possesses, CFO John Felton addresses AI infrastructure head Prasad Kalyanaraman: “You’ve made our company have negative free cash flow, congratulations.” According to the publication, Amazon’s capital expenditures on AI construction reach $220 billion in 2026, and the company has fallen into negative free cash flow for the first time in a long time. At the same meeting, Kalyanaraman confirmed that the machine learning and AI business is “massively negative on free cash flow” and will remain so for “another couple of years.” Meanwhile, demand for AWS capacity is confirmed by operational metrics: revenue in the second quarter grew by 37%, the highest in 18 quarters, the portfolio of signed contracts reached $496 billion, and 44 GW of data center capacity is under development across 16 US states.

Context

Free cash flow is the money remaining with a company after operating expenses and capital investments, so large-scale data center construction directly reduces this metric. Major cloud companies have historically tried to spend capital at the last responsible moment, purchasing equipment shortly before launching capacity, but in the AI race, this scheme has stopped working. The bottleneck has shifted from chip production to electricity, construction, and labor, and equipment now has to be purchased in advance and kept in warehouses, freezing billions of dollars. According to Business Insider, since August 2026, Amazon has been receiving approximately $5 billion worth of Nvidia GB300 accelerator racks per month as part of an internal Early Rack Receipt process and expects around $3 billion per month in the future, even though the data centers for these racks are not yet ready. The company tracks power grid risks through an internal dashboard called Secured Power at Risk, is building housing for electricians and plumbers in West Texas, and even includes shipping containers in its expense items. In other words, the physical world has become the main limiter for AI scaling.

Why this matters for the industry

For the industry, the material shows that the financial model of the largest cloud provider has been consciously restructured for a multi-year AI race: leadership is willing to fund negative cash flow for the sake of a future computing base. According to Business Insider, Google, Meta, and Microsoft together are building less than half of the 44 GW that Amazon is planning to deploy, meaning that if the projects are realized, the company is laying a structural advantage in available capacity for years to come. However, new capacity is being pre-occupied by Amazon’s internal AI workloads, so availability of compute for independent teams may not improve immediately: deployment will be distributed first among contracts from the already signed portfolio. The horizon allows for two scenarios: an oversupply of capacity with pressure on inference prices, which would be a gift to product teams, or market consolidation and price increases. In both cases, teams that calculate compute costs in advance and know how to shift load accordingly will benefit.

Why this matters for users

For teams building products on cloud AI, this brings no direct changes: the material contains no new APIs or pricing announcements. The practical signal is that the shortage is forming on the infrastructure side, so it makes sense to re-check GPU quotas and backup contracts with compute providers now, build regional capacity limitation risks into the architecture, and strengthen inference cost optimization: caching, cheaper models for background tasks, and budgets for agents. For those following AI infrastructure, the material offers a rare look inside how scaling hits reality: from controlling power grids to warehouses with equipment waiting for ready sites. This cross-section is useful for forecasting availability and compute prices, and one should expect regional unevenness in capacity deployment, against which the emergence of third-party energy risk monitoring tools is likely.

What is currently unknown / limitations

The material relies entirely on a single Business Insider publication, based on a leaked internal audio recording of a meeting from September 2026; there is no independent verification or official comment from Amazon at the time of publication, and the discussion in the Hacker News thread consists of only 4 points and 1 comment. The composition and terms of the $496 billion contracts are not visible from the text, so the large portfolio cannot automatically be considered proof of sustainable long-term demand. AWS internal tools, including the Secured Power at Risk dashboard and the Early Rack Receipt process, do not have a public API and are described only according to the leak. Finally, the comparison of Amazon’s 44 GW with the combined capacity of Google, Meta, and Microsoft is provided according to the same publication and requires independent verification.

Sources

Author

Look at AI, editorial team