Industry executive Gavin Baker forecasts that the AI compute shortage will persist through 2028, with data center payback periods extending to nine months.

Venture investor Gavin Baker argues that the AI industry is not experiencing an oversupply bubble but rather a severe, self-inflicted compute shortage that will persist through 2028. Speaking on the a16z Podcast with host David George, Baker presents a case built on extraordinary infrastructure economics: gigawatt-scale data centers are generating sub-one-year paybacks, with Nebius disclosing nine-to-ten-month returns on deployed capital. Demand is accelerating across every frontier lab, open-source project, and enterprise application, while supply remains constrained by power availability, land acquisition, and regulatory friction. Fewer than 10 million heavy AI users currently support roughly $80 billion in annualized lab revenue, leaving 1.5 billion knowledge workers yet to adopt the technology. The conversation frames Nvidia as the "central bank of AI," citing CEO Jensen Huang’s eight-to-nine-chip product lineup, supply chain lockup, and alignment with ecosystem fragmentation as factors that uniquely position the company. Beneath the technical narrative lies a political dimension: regulatory obstruction, which George attributes partly to a coordinated campaign, could create compute inequality where only large corporations and wealthy individuals can afford frontier intelligence. Concluding with SpaceX’s orbital data center plans and the economics of Starship reusability under Elon Musk, the episode paints a picture of an industry whose technical momentum is extraordinary but whose ability to tell a compelling story to the American public is failing.

In late August 2026, public AI stocks were in a drawdown, but Baker observed something entirely different in the underlying businesses. “My standard question is, can you tell me one quantitative data point in your business that’s getting worse? Just one,” the technology investor said on the podcast. “It’s at least in July and August, I haven’t been able to find a single person.” Speaking with George, Baker delivered a striking thesis: the AI sector is suffering from a severe compute shortage, not an oversupply bubble. Demand is accelerating across frontier labs, open-source initiatives, and enterprise applications, yet the market has missed it because equities are falling. “Public stocks have kind of fallen out of bed over the last two months,” he noted. “You can drown crossing a river that’s on average 2 ft deep.”

The disconnect between market sentiment and business reality is stark. Baker reports that OpenAI has clearly accelerated, open-source models have accelerated even more, and Grok—particularly following the launch of Grokbot—has seen dramatic usage growth. Yet the stock market is pricing in a slowdown. The scale of untapped demand forms the core of the bull case. Baker estimates that fewer than 10 million heavy-paying users, primarily developers and enterprises extracting real value, support roughly $80 billion in annualized revenue across major AI labs. The global knowledge workforce comprises 1.5 billion people, and adoption diffusion is nowhere near complete.

Baker points to a power-law distribution within corporate environments: the highest-spending engineers consume ten to one hundred times more tokens than the median engineer. Within his own portfolio, the most AI-native companies allocate high single digits to over 10% of human compensation costs toward tokens, while traditional enterprises adopting the technology effectively spend around 1%. The demand trajectory is explosive. George cites internal a16z data showing token consumption grew 100-fold from March to August 2026. This represents not incremental adoption, but a step change.

A central tension for frontier labs involves the deliberate choice between allocating compute to revenue-generating inference versus research and training. George walks through the math: if a lab possesses 10 gigawatts of power and allocates 8 to inference, monetizing at $60 billion per gigawatt annually yields $480 billion in revenue—a one-year payback on a revenue basis. However, a research breakthrough could flip that allocation, dropping revenue from $480 billion to $120 billion overnight. This creates structural volatility absent in prior tech eras. “There was no massive trade-off they had to make in terms of the cost or infrastructure to serve revenue side,” George said, contrasting AI with the internet era. “Like they were totally separate.”

The strategic conclusion is clear: labs will not prioritize free cash flow. “I don’t think any of them are going to be that focused on generating free cash flow,” Baker said. “They’re going to generate a lot of operating cash flow and then they’ll use that to buy a lot of GPUs.” Historical evidence supports aggressive spending. George cites Microsoft CEO Satya Nadella “blinking” on capital expenditures last year and regretting it, and notes that Anthropic CEO Dario Amodei’s publicly conservative spending stance allowed competitors like OpenAI and SpaceX to surge ahead through aggressive investment. Baker frames the dilemma bluntly: “If you don’t spend enough you could lose a lot of shares. But if you spend too much, you could go bankrupt. And like those are both bad things, but bankruptcy is worse than losing shares.”

The economics of building AI infrastructure are currently extraordinary. Disclosures from Nebius and CoreWeave indicate a nine-to-ten-month payback period for gigawatt-scale data centers. The financing math is compelling: a $50 billion build requires only a $15 billion equity check because customers prepay 50% to 60% of capacity, and the remaining capital can be financed at low cost by sophisticated institutions. Baker emphasizes that this is not circular financing. “I know a lot of smart people who work at Blackstone and KKR and Apollo and they’re the ones that are financing it at a relatively low cost,” he said. The useful lives of GPUs keep extending as models improve, and monetization rates per gigawatt continue rising, making the true equity payback potentially “way inside of a year.” George frames this as a career rarity: “In my career as an investor there haven’t been that many opportunities where you have companies that could deploy tens hundreds of billions of dollars and get sub one-year paybacks.”

The demand side reinforces the supply-side economics. If demand continues to outpace buildout—which George believes is likely through 2028 given regulatory delays—prices for frontier tokens could rise dramatically. Baker cites a provocative point from Dario Amodei suggesting token costs could increase tenfold. The consequence of compute inequality, driven by data center opposition activists, would be a world where only large corporations and wealthy individuals can afford frontier intelligence. Both speakers express frustration with organized opposition to data centers in the United States. George attributes part of this resistance to a “CCP funded campaign” laundered through TikTok. However, the core problem is narrative: the industry has failed to tell a compelling story about tangible benefits to ordinary Americans. “You’re opposed to data centers. Well, you know what? It’s probably the best thing that has ever happened to working-class Americans,” George said. “We are re-industrializing America and it’s awesome.”

The data supports this perspective. When a data center operates behind the meter, it transforms a town—tax revenue doesn’t double, it increases tenfold, revitalizing struggling small communities. George cites Loudoun County, Virginia, which boasts the highest per-capita income of any U.S. county alongside the highest density of data centers. Water consumption concerns are “totally debunked,” and natural gas serves as a relatively clean fuel source. The “stay ahead of China” argument has been “correct but ineffective” because it remains too abstract. Baker argues the burden of proof is shifting: the industry must deliver tangible everyday benefits beyond answering questions or substituting for search engines. He praises Meta for doing the best job of communicating this value, referencing former COO Sheryl Sandberg’s practice of highlighting ten to fifteen specific small businesses transformed by Meta’s advertising products during earnings calls. The AI industry—including NVIDIA, AMD, Broadcom, and all the labs—should adopt the same approach, particularly as competing silicon architectures like Trainium enter the fray.

The speakers converge on a future that is decidedly not a two-horse race between OpenAI and Anthropic. Baker describes it as an “and thing, not an or thing”: frontier models will perform well, N-minus-one models will perform well, open-source systems will thrive, and a host of application companies will succeed. “This is not an or thing, it’s an and thing,” he said. “Frontier is going to work really well, N-minus-one models are going to work really well, open source will work well, and a host of application companies will thrive.”