Summary

David George’s base case is that AI infrastructure is being financed by companies strong enough to absorb overbuild while the cost-performance curve compounds in application developers’ favor. Annualizing the latest quarter puts big-tech capex near $400 billion, “most of that” for AI infrastructure and data centers; model-access costs fell more than 99% in two years while frontier capability doubled every seven months. He expects AI to become “like electricity or Wi-Fi,” with large tech companies carrying much of the substrate cost.
The dot-com analogy breaks, in George’s view, because today’s capacity sits atop internet and cloud distribution and already has usage at global scale. ChatGPT reached 365 billion searches in two years versus Google’s 11, while George estimates 1.5–2 billion active AI users across products; the builders and tenants are stronger, though leverage routed through banks, private debt, and insurers remains worth watching. “It’s built on the back of the previous technology cycles.”
The addressable value pool is labor, not merely software: US software spend is about 1% of GDP versus white-collar payroll near 20%. George expects AI to exceed the roughly $10 trillion of value created by mobile and cloud, with perhaps 90% of value accruing to customers and 10% to vendors—still enough for enormous market caps. If completed work remains hard to measure and price, competitive forces will leave even more surplus with users.
Consumer AI may surprise on price before it surprises on reach. ChatGPT was described as having more than one billion monthly active users and 30–40 million paying users, versus perhaps two billion AI users overall; India pricing near $3–4 a month coexists with US premium products at $200–300. George thinks the P in P×Q has substantial runway because “there’s way more upside to monetize the base than there is risk of price pressure.”
Kha and George’s bottleneck view is that chip and infrastructure capacity should scale, but energy is likely the limiting factor for the next five years and cooling follows behind it. Kha points to nuclear, expects Three Mile Island to get powered back up, and cites West Texas natural gas; xAI’s speedrun required buying backup generators across a multistate region and labor off other projects. Kha’s next constraint is cooling the buildout “without boiling our oceans” or melting the chips.
For AI applications, George would accept temporarily weaker gross margins—but not weak product love. The underwriting hierarchy is 90% or more gross retention and easy customer acquisition ahead of current margin, conditional on multiple model suppliers driving inputs lower; GPT-5, Anthropic, and Gemini were cited as competitive pressure. Consumer products can be sticky, while raw developer APIs are “not very sticky” because switching can be one API call.
The investable high-growth frontier has migrated into private markets, making access and liquidity—not just selection—core return variables. Billion-dollar private companies total roughly $3.5 trillion versus $500 billion ten years ago, companies now remain private for about 14 years, and only around 5% of public software and internet names forecast growth above 25% for the next 12 months. a16z’s approach pairs “undeniable momentum” with unusually early bets on only the strongest research teams.
Incumbent software is vulnerable only where a startup can combine three breaks at once: reimagined UI/UX, a new data layer, and disruptive pricing. Salesforce is George’s example of an uninspiring front end attached to a sticky database; AI can shift software from keeping records to doing work, but he has not yet seen the killer dethroning idea. Near-term opportunities sit around systems of record rather than in wholesale replacement.
Beyond AI infrastructure and applications, George expects American Dynamism to be the next-largest area, with some AI-enabled health activity and crypto pursued alongside the crypto team. Stablecoin enablement could become more significant if that market takes off. The portfolio follows best ideas rather than a quota for new investments versus follow-ons, and George says a16z’s edge also comes from early-stage access plus market and product insights.

Deep dive

1. Large tech companies are underwriting an AI utility layer

George opens from market structure: five—and sometimes six—of the most valuable companies are US technology companies, with seven or eight in the top ten. “Technology has swallowed the whole market,” while AI is making companies and investment amounts larger faster than a16z has previously seen.
Annualizing the latest big-tech quarter produces roughly $400 billion of capex, with “most of that” going into AI infrastructure and data centers. The favorable supply-side feature is who bears the burden: Google, Facebook, Amazon, and Microsoft can absorb excess capacity in a way that the companies building previous infrastructure cycles could not.
Simultaneously, model-access costs have fallen more than 99% in two years—roughly a 100-fold decline—while frontier capabilities have doubled every seven months. George says the decline is greater than the decrease predicted by Moore’s law, giving applications improving outputs and falling inputs at the same time.
The economic comparison is deliberately expansive: mobile plus cloud created roughly $10 trillion of market value, but US software spend represents only about 1% of GDP against white-collar payroll near 20%. George’s “rule of thumb” is that customers capture 90% of new value and vendors 10%; Apple and Google show that enormous consumer surplus can coexist with extraordinary businesses.

2. Existing distribution makes this buildout unlike broadband

Kha’s pushback—worth keeping—is that the capex charts resemble the early-2000s broadband buildout and its painful glut. George says the world eventually “grew into it,” but stresses that less-strong companies were building out that cycle; today’s hyperscalers and tenants are stronger, though “the thing to watch” is leverage.
Private capital does not remove systemic linkages: George notes that banks fund private-debt companies, while insurance companies increasingly sit behind them and may be backdoor funding the buildout. He nonetheless regards the identity of the builders and tenants as “a really good sign for the stability of the buildout,” rather than claiming financing risk has disappeared.
Demand arrived on a radically compressed clock. ChatGPT reportedly needed two years to reach 365 billion searches, versus 11 years for Google, because AI inherits global internet access, smartphones, and cloud infrastructure instead of waiting for a new network or hardware device. That is “immediate global distribution.”
George estimates that more than one billion people use ChatGPT monthly, perhaps another billion have tried it, and 1.5–2 billion people actively use AI products in some form. The exact totals are hedged, but the conclusion is categorical: “The speed at which they got to distribution is unlike anything we’ve seen before.”

3. AI can price-discriminate where prior platforms could not

Google, Facebook, and Apple historically lacked a clean way to charge each consumer according to willingness to pay. AI is already segmenting: OpenAI’s India product was cited at roughly $3–4 per month, while high-end US subscriptions at $200–300 were described as “flying off the shelves.”
George’s shopping example carries the commerce argument: a deep-research product compared detailed specifications and year-over-year value for his son’s baseball bat, producing “extraordinary answers” without the process of typing into Google, clicking around, and seeing seven sponsored links before organic results. That experience could support advertising or affiliate-like monetization while reducing referral traffic to websites and companies.
The discussion puts OpenAI at roughly 30–40 million paying users, with perhaps another 10 million across competitors, against around two billion AI users overall. Daily ChatGPT users already spend 28–29 minutes in the product, compared with roughly 50 minutes for Instagram and 70 for TikTok—evidence of meaningful attention before broad monetization.
On concern over billion-dollar monthly burn, George says most of the burn comes from research and development and future investments. Free competitors have not materially pressured the consumer business over the past 12 months, though he says that might change; meanwhile, consumer familiarity is stickier than raw developer APIs, which applications can replace with “an API call” when a better coding model appears.

4. Energy is the five-year bottleneck; cooling comes next

Kha says that, under current means of energy production, energy is a bottleneck. She is most optimistic about nuclear, expects Three Mile Island to get powered back up, mentions data centers locating near nuclear plants, and points to West Texas natural gas as another way to power large training clusters efficiently. George agrees that the bottleneck will shift once this problem is solved.
Kha says xAI erected what was then the largest data center in roughly one-quarter the normal time, but only through “crazy unnatural things”: acquiring backup generators across a multistate region and buying labor away from other projects.
Kha’s view is that chip and infrastructure production capacity will typically scale toward demand after periods of dislocation, leaving energy as the probable bottleneck over the next five years. She then identifies cooling as the next major constraint: the industry must innovate “without boiling our oceans,” while George adds that it must avoid making the chips melt down.

5. Product love outranks today’s gross margin

George welcomes the unusually wide range of AI outcomes: investors must decide not only market size, but business quality, market power, which layer wins, and whether value accrues even to a winner. That variance makes the period more difficult to underwrite, but also creates the possibility of more differentiated investment returns.
His two preferred top-line tests are gross retention and ease of customer acquisition. If 100 customers begin the period, he wants 90% or more of their dollar value to remain, ideally with expanding usage; equally important is organic demand and high willingness to pay relative to sales and marketing cost—products being “pulled off the shelves.”
Gross-margin leniency rests on a condition, not faith: multiple near-par model providers must keep pressuring input costs downward. George expects the 100-fold decline to continue, perhaps more slowly, and cites GPT-5 as a credible alternative to Anthropic plus improving Gemini coding models from Google.
The portfolio will not accept “a bunch of companies with zero gross margins.” But compared with mature SaaS, George gives today’s margins more benefit of the doubt when retention and acquisition are exceptional, because cheaper inputs and better models could simultaneously raise product value, deepen stickiness, and improve unit economics without higher prices.

6. Workflow depth—not model access—creates application durability

Medical scribes, customer support, and high-end financial analysis look relatively sticky because the model becomes surrounded by integrations, company-specific rules, workflow sequencing, and enterprise capabilities. Brands also encode a preferred interaction style, making replacement more consequential than merely swapping one underlying model for another.
Experimental internal-tool creation and low-end website prototyping look less durable. George expects the market to bifurcate between products used to sketch prototypes and tools trusted to build and deploy real applications; it remains “TBD” who owns each segment, and companies are unlikely to “vibe-code up their Salesforce.com.”
The hoped-for pricing progression runs from perpetual licenses, to SaaS seats, to cloud consumption, and ultimately to monetizing the replacement of tasks humans do. Customer support is furthest along because a task can be definitively resolved, but elsewhere customers still prefer familiar seat or usage pricing and objective task completion remains difficult to measure.
George is “low conviction” that all software will adopt a new business model within five years. Without credible outcome measurement, vendors cannot capture the full value of replaced labor: like the steam engine, AI will be priced under competition and a return-on-capital constraint, leaving much of the productivity surplus with customers.

7. Private markets now own most high-growth discovery

The old “triple-triple-double-double” growth benchmark now looks modest. George says leading companies have reached $10 million and then $100 million of revenue perhaps four times faster, so a16z compares new applications with Cursor, Decagon, Abridge, and ElevenLabs—not with historical trajectories from Shopify or DocuSign.
Companies once went public five to ten years after formation; the figure is now around 14 years and still lengthening. Private companies valued above $1 billion collectively represent roughly $3.5 trillion, or about 10–12% of the Nasdaq, versus approximately $500 billion ten years ago—a sevenfold expansion.
Something like 5% of public software and internet companies forecast growth of 25% or more over the next 12 months. Figma and other strong businesses may still list, but George does not expect the structural migration of high growth into private markets to reverse soon; accordingly, the opportunity cost of owning slower public names is unusually high.
George says the fund makes very few public-company investments because the bar is extraordinarily high and the private opportunity set is growing faster. Longer private lives create a real DPI tension rather than a free option: he favors regular tender offers that give employees liquidity and help private companies compete with quarterly vesting public-company RSUs. a16z wants what is strategically best for each company, but ultimately must monetize investments and does not dismiss the hold-period problem.

8. Access and asymmetry define the portfolio

The AI strategy has two buckets: companies with “undeniable momentum,” such as Cursor, Decagon, ElevenLabs, and Abridge, and unusually early growth investments in what George calls the top five teams in the world. Early relationships produce “ball control” in later rounds; he says a16z supplied xAI’s first outside money beyond Elon.
George distinguishes the return profiles rather than applying one to the whole portfolio. For a recent Databricks opportunity, he described 2x as relatively safe and 3–4x as a confidence range, while saying the fund does not want a portfolio made entirely of such investments. For champion companies, he described perhaps 2x in a severe downside, with the possibility of 5x over five years but not necessarily 10x.
Research-team bets have wider business outcomes but potential capital asymmetry because exceptional talent remains valuable even when the original company plan fails. George will not set a quota for high-variance research teams: “There’s not another Ilya floating around in the AI market.” Those investments happen reactively when rare people emerge, with the AI infrastructure team helping assess the teams and the research ideas they are pursuing.
Beyond AI infrastructure and applications, George expects American Dynamism to be the next-largest area, with some AI-enabled health activity. Crypto investments are pursued alongside Chris and the crypto team when opportunities fit the growth fund, with stablecoin enablement a particularly exciting area that could receive more attention if the market takes off. The portfolio follows best ideas rather than a quota for new investments versus follow-ons.
Public-software disruption requires three ingredients together: reimagined UI/UX that “does things for you,” a new data substrate incorporating unstructured information, and business-model innovation against seat pricing. George has not found the definitive Salesforce killer; a16z instead sees openings around systems of record, informed by early-stage relationships that precede roughly 80% of its new non-early-stage investments.
George says a16z’s other major source of alpha is market and product insight. The close connection to the early-stage teams helps the growth team see emerging companies and categories earlier while allowing a relatively small team to cover substantial ground.

Vol.205 Chen Qigang, the Musician, and The Hidden Mountains and Rivers: Talking Creative Work with Director Guo Xufeng Legora CEO, Max Junestrand: $7M ARR in a Day | Harvey vs Legora: Is a Legal AI Winner Takes All?