Power shortages and debt threaten to halt AI growth

Physical walls are stopping the artificial intelligence boom. On August 25, 2026, tech analysts and investors began charting a sharp collision between relentless token demand and the hard limits of power grids, memory chips, and global capital.

On the Dwarkesh Podcast on August 25, 2026, SemiAnalysis founder Dylan Patel projected that Anthropic and OpenAI will capture half of all new global compute capacity by late 2027. Serving modern reasoning models now yields up to $50 million per megawatt in revenue, allowing frontier labs to outbid legacy enterprises. To fund this expansion, infrastructure spending will reach $11 trillion by 2029. Patel warned that hyperscalers will flood credit markets with high-yield bonds, raising interest rates and choking out traditional corporate borrowing.

"Capital constraints will hit the market faster than physical supply bottlenecks. Financial markets will choke on AI debt long before silicon supply runs out."

- Dylan Patel, Dwarkesh Podcast

That financial model faces immediate skepticism from open-source advocates. Speaking on BTC Sessions on August 25, 2026, investor Jeff Booth argued that massive infrastructure spending is structurally unpayable. Because software code generation trends toward zero marginal cost, open-source models rapidly replicate proprietary breakthroughs at zero incremental cost to users. Booth compared the spending surge to the dot-com bubble, warning that free models will strip away subscription revenues and destroy expected investment returns.

Three days later, Andreessen Horowitz general partner Ben Horowitz countered that software economics no longer apply to frontier models on The a16z Show. Compute supply across chips, memory, and high-voltage electrical transformers is booked out through 2028. Silicon memory suppliers face a three-year backlog, forcing buyers into multi-day auctions for secondhand hardware. Data center rack power density has soared from 10 kilowatts to 250 kilowatts, but only 2 percent of American electricians are certified to install high-voltage direct current systems.

On the same August 28 broadcast, general partner Martin Casado explained why capital continues to flood the space despite operational bottlenecks. Historically, adding software engineers to a late project slowed progress. In AI development, capital converts directly into raw compute capacity, allowing companies to bypass human engineering cycles by spending billions on massive server clusters. With hyperscale capital expenditure set to hit $1 trillion next year, the bottleneck has moved entirely from software design to factory production rates.

By August 30, 2026, Andreessen Horowitz managing partner Jen Kha revealed on The a16z Show that hardware pitches jumped from under 5 percent to over 20 percent of incoming deals, prompting the firm to launch its $1.1 billion Machine Age Fund. Kha warned that domestic regulatory delays and environmental opposition are driving infrastructure projects overseas. While foreign governments in South Korea and El Salvador build AI public utilities, American startups are moving data clusters to Mexico and Australia to bypass domestic power permitting backlogs.

The physical constraints culminated on August 31, 2026, when investor Gavin Baker argued on The a16z Show that the market faces a severe compute deficit through 2028. Rather than facing an overinvestment bubble, cloud providers see capital payback periods under ten months, driven by software power users who consume 100 times more tokens than median workers. To bypass land, copper, and electrical grid tie-ins that eat up $15 billion per terrestrial gigawatt, Baker highlighted plans with SpaceX and Nvidia to deploy orbital data center racks using Starship launches by late 2027.

"The threat isn't overcapacity. It's compute inequality where only well-capitalized firms can afford intelligence."

- Gavin Baker, The a16z Show

The compute bottleneck is redrawing global technology.