Gutenberg Lab — Morning Paper

AI Demand Outrunning Compute Supply — Sub-One-Year Paybacks Signal Structural Undersupply

Featuring

David George (General Partner, a16z) × Gavin Baker (Investor, Atreides; formerly Fidelity)

Bottom line

AI compute demand is massively outrunning supply, with GPU infrastructure achieving sub-one-year payback periods and undersupply projected through 2028 rather than the overbuild many fear.

3-Line Summary

Investor Gavin Baker spent summer 2026 asking AI leaders for any worsening metric and found none—OpenAI, open source, and Grok all accelerated in July-August even as public AI stocks fell 20-30%. The current $180 billion revenue run rate monetizes maybe 30 million heavy users versus 1.5 billion knowledge workers—demand has barely started. Physical constraints (copper, wafer capacity, power) prevent overbuild, making undersupply through 2028 the real risk, not oversupply.

3 Key Points

1. Sub-one-year paybacks at unprecedented scale: Nebius and CoreWeave disclosures show 9-10 month paybacks for $50B/GW data centers—customers pay 50-60% upfront ($25-30B), remainder monetized in spot market at $2-3/hour. Blackstone, KKR, Apollo underwriting these deals. Baker: "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." This isn't speculation—it's happening now with sophisticated debt financing.

2. Demand side barely started—30M users vs. 1.5B knowledge workers: Current $180B revenue run rate monetizes maybe 30M heavy users (Baker: "I might take the under"). Within companies, highest-spending engineers spend 10-100x more than median. Atreides' token consumption up 100x March-August 2026, may 10-20x again in one month with GrokBot Enterprise (two users). George: "There's 1.5 billion knowledge workers. We're nowhere on the demand side, and we're massively supply constrained."

3. Physical constraints prevent overbuild—risk is undersupply, not bubble: Copper, wafer capacity, power, cooling—all hitting fundamental limits. Jensen locked up 70-80% of fab capacity, DRAM, NAND, lasers, capacitors, rack components. Regulatory delays mean no capacity available through 2028 with all forecast builds. Baker: "We're in this acute shortage with several million people driving a crazy global compute shortage. What happens when that's 500 million? How many copper mines do we need to build?" Consequence: compute inequality where only wealthy can afford intelligence.

Editorial Perspective

This is a classic platform-transition supply-demand mismatch, but with a critical difference from past cycles: physical constraints are preventing overbuild. During the internet boom, capital raced ahead and overbuilt fiber infrastructure. This time, copper, power, and fab capacity—real-world physical assets—are acting as natural governors. The result: the risk isn't oversupply but undersupply, where only wealthy users can afford access to frontier intelligence. The analogy isn't the dot-com bubble but the early days of electrification, when power generation couldn't keep pace with demand and access became a class divider. The difference is that intelligence, unlike electricity, compounds—those who get it first pull further ahead. That's why Baker's framing matters: this isn't "when does the bubble pop?" but "who gets left behind in compute inequality?"

Source

This Week in Startups, "Gavin Baker: Why AI Demand Is Outrunning Compute Supply" (August 31, 2026)

https://a16z.simplecast.com/episodes/gavin-baker-why-ai-demand-is-outrunning-compute-supply-z97oFoOG

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Texas battery market faces 84% revenue collapse — will data center demand rescue merchant economics?

Speakers

Shayle Kann (Partner, Energy Impact Partners) × Brent Vermillion (US Market Lead, Moto Energy)

Bottom line

Texas batteries suffered an 84% revenue collapse due to overbuild, but explosive data center demand growth could restore volatility by 2029–2030 — unless behind-the-meter generation suppresses price spikes and leaves merchant batteries stranded.

3-line summary

• ERCOT (Texas grid) battery capacity grew 100x from 200 MW (2020) to 16.5 GW (mid-2026), but average revenue fell 84% from ~$200/kW (2023) to ~$30/kW (2025), well below the ~$100/kW needed to finance a 2-hour system
• The collapse is driven primarily by overbuild, not mild weather — ancillary services saturated, forcing batteries into energy arbitrage, which they then cannibalized by competing to discharge and flattening the spreads they were chasing
• Developers bet that 400–500 GW of data center interconnection requests (inflated) will transform the market if even 10% materialize, but 13–15 GW of batteries withdrew from the queue in 2025 after signing interconnection agreements — historically a near-guarantee of completion, making this the first clear supply-side slowdown signal

Three key points

1. Storm Fern (Jan 2026) revealed a new risk: batteries depleting early, then prices spiking when the fleet is empty. Batteries discharged into an anticipated morning peak that didn't materialize. The next morning, only ~1/3 of the fleet had state of charge, and real-time prices hit 4 figures. Vermillion: "Battery state of charge is such a key variable for optimizers now." [00:18:39, 00:19:59]

2. Duration is lengthening from 1 hour to 2 hours. ERCOT batteries averaged ~1 hour a few years ago; now nearly 2 hours, with some 2.5–4 hour projects. Drivers: shift to energy arbitrage and the need to capture secondary price spikes when shorter batteries are depleted. [00:16:02, 00:16:28]

3. PJM is repeating ERCOT's early cycle. PJM's regulation market redesign created "pay back in one year" prices, but the market is only 600–700 MW deep. PJM has ~500 MW of batteries now; at 1–2 GW, prices should compress. [00:36:03]

Editorial perspective

The battery overbuild mirrors the early internet boom — capital rushed in on technological promise, then collapsed waiting for demand to catch up. ERCOT's bet is the classic market cycle theory: "low prices cure low prices." But if data centers bring behind-the-meter generation, volatility itself disappears. This market is the global test case for merchant battery economics in energy-only markets, and the outcome will shape deployment strategies across North America.

Source: Catalyst "ERCOT batteries: boom, bust, or rebound?" (October 1, 2026)

https://www.latitudemedia.com/podcasts/catalyst

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Survival Strategies for Women in Crypto: A Decade of Experience Condensed

Speakers

KK Katherine (General Counsel, Chainlink Labs) × Jesse (former DOJ crypto prosecutor, now Rivet Capital) × V (former SEC enforcement attorney, formerly Worldcoin)

Bottom line

Surviving as a woman in crypto requires three defensive lines: deep technical understanding, intellectual independence, and relationships outside the industry.

3-Line Summary

• The final episode of "DEX in the City," a crypto podcast, featured three women hosts—all senior legal/policy professionals who transitioned from government or traditional finance—sharing survival strategies accumulated from 2016 to 2026
• Despite the industry's growth, they acknowledge "the situation for women in crypto has not gotten any better this year" (Jesse), offering advice across three axes: credibility-building, mental health, and resisting conformist pressure
• Crypto is no longer a separate industry but "everything," requiring cross-disciplinary fluency in regulators, markets, and technology; yet women face a structural disadvantage where technical knowledge is a credibility tax men may not pay equally

3 Key Points

1. Technical understanding as a credibility tax: V (former SEC enforcement attorney) states bluntly, "People will make assumptions about you as a woman...you need to learn the technology." Not coding necessarily, but "how the product works, who has control, how the assets move, how the smart contracts operate, where all the risks are." In a male-dominated industry, this deep understanding is a prerequisite for credibility—more so in crypto than other industries, she argues

2. Conformist culture born from existential threat: V analyzes crypto's "very conformist culture" as a product of survival mode: "For so long it was us against the world. We didn't even know if we would survive as an industry for many years. So this kind of solidarity developed." Existential threat created intellectual homogeneity and raised the social cost of dissent. Jesse, whose focus on illicit finance and victim protection often conflicts with industry orthodoxy, advises: "Don't lose yourself. There is space for you in tech and crypto if you want to be there"

3. Identity fusion as a vulnerability: V recounts breaking down mid-presentation at a Senate meeting shortly after FTX collapsed, requiring staffers to hand her a tissue. "I'd attached part of my identity to the belief that this industry was worth defending. So FTX felt like personal betrayal even though it had nothing to do with my work." For those who left stable careers for crypto, defending the industry became tied to defending their life choices. "You can care a lot about this work without letting it become your whole identity"

Editorial Perspective

Crypto's maturation has been a war of legitimacy, not innovation—a "bullheaded push," in Jesse's words. The core technology (blockchain, tokenization, tokens moving without government control) hasn't fundamentally changed since Bitcoin. What changed is that the FBI and banks that once "laughed us all out" are now curious or leaning in. As a historical record of this decade, the episode documents not technical achievement but human cost. Women expected to be "the adult in the room," paying technical knowledge as a credibility tax, and deliberately maintaining relationships outside the industry to preserve mental health—these structural issues persist even as crypto becomes "everything." The industry's growth has not yet solved the problems it created for those who built it.

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --