Keith Rabois: AI app-layer companies must have positive gross margins, and the 'deal guy era' echoes the dot-com bubble

Summary

Keith Rabois is watching the current venture environment with visible unease. Valuations have 'an extra zero,' and he notes his own historical ceiling for entry price was $86 million — a figure that now looks quaint. He invested in Airbnb at a $3.5 million post-money valuation and DoorDash at $8–10 million post. That world is gone, and he believes a meaningful share of capital being deployed today will be lost.

AI App Layer: Margins Are Non-Negotiable

Rabois has put two-thirds to three-quarters of his investments over the last 18 months into AI, predominantly at the application layer. His standard is firm: every application-layer company must carry positive gross margins from day one, and not modestly positive. He draws a hard line between infrastructure and frontier model builders — where losses are expected and rational — and workflow software companies that are simply calling APIs and wrapping them in a UI. Companies like Harvey in legal AI, he argues, have no defensible reason to be burning capital at scale given that compute costs should be passable as margin.

His portfolio bet in legal AI is Spellbook, which he describes as serving corporate legal departments at companies including eBay, helping them enforce firm-specific legal preferences in software rather than through billable hours. Corporate legal buyers, unlike law firms, are incentivized to spend less and move faster, which creates natural pull for automation tools.

When evaluating AI revenue, Rabois looks past the contract. He calls customers directly to assess whether business goals are actually being met, examines whether revenue is pilot or permanent, and weights engagement metrics — MAUs, DAUs — as heavily as revenue figures. A $5 million enterprise contract that costs $10 million to deliver is, in his framing, not a business.

ChatGPT Has a Real Ceiling — If OpenAI Stays Focused

Rabois rates ChatGPT as still the fastest-growing consumer product in 15 years and believes it could be a $4 trillion business if OpenAI executes with discipline. The risk, in his view, is distraction. Pursuing AGI research, hardware, consumer electronics, and science simultaneously competes for the same finite pool of talent and executive bandwidth. He tested GPT-4o thinking mode against Gemini on an identical book proposal prompt and found ChatGPT retained what he described as a 'spark of ingenuity' and human personality that Gemini did not replicate.

On hardware, he thinks Sam Altman likely has to make a device move, not because it is obviously right, but because the downside of being caught flat-footed by a competitor that integrates AI into consumer hardware first is too severe.

Apple Is Behind, But the Lock-In Is Real

Apple is the one major tech company Rabois believes is not paying adequate attention to AI. He contrasts this with the rest of the Magnificent Seven, which he considers genuinely engaged. He gives Apple's hardware — including vertical integration, supply chain, and battery — continued credit, but puts a 10-year horizon on a credible AI-native device from a competitor unseating it. Near-term, he sees the moat holding. Longer term, he would not bet against disruption.

On Google, he acknowledges the shift in assessment post-ChatGPT. What once looked like a sustaining innovation challenge now looks potentially disruptive. But Google's ability to monetize gives it an R&D runway that OpenAI cannot match without continued fundraising.

The 'Deal Guy Era' Echoes the Dot-Com Bubble

Rabois flagged what he sees as a culturally dangerous trend: the re-emergence of business-development-first operators in Silicon Valley. He traces the archetype directly to the 1996–2000 Internet bubble, when 'biz dev' professionals ran deal structures that inflated valuations without underlying product value. Post-collapse, the label became a career liability and disappeared for roughly 25 years. Its return, in his reading, is a meaningful warning sign that the cycle may be repeating.

His framework remains 'bold, early, impactful' — first institutional capital, clear value proposition, and a product that measurably moves the P&L of the customer. He points to an early enterprise investment where Walmart spent $5 billion and a telecom spent roughly $10 million specifically because the technology addressed top-three company priorities — in telecom's case, subscriber churn, which he describes as the only metric that actually matters in that industry.

Consumer AI Commerce: Conversion Is High, Volume Is Not Yet

Rabois is watching AI-driven commerce closely. A Shopify brand doing $200 million in revenue reported that visitors arriving from ChatGPT convert at 12% — meaningfully above typical e-commerce benchmarks. He attributes this to the authoritative nature of LLM recommendations relative to a standard search blue link. His concern is that volume at that entry point remains low, and the catalyst for scaling it is not yet clear.

His portfolio includes ProFound, a company that tracks how AI platforms including Gemini and ChatGPT surface brands and products — infrastructure he views as critical as AI-driven discovery grows.

On the consumer side more broadly, Rabois is skeptical about the scale of breakthrough. He believes ChatGPT and OpenAI will be a generational consumer company, but is not confident about who else qualifies. The constraint is not technology but time: consumers have 24 hours in a day, and every new app competes with family, friends, and existing habits. The bar for displacement keeps rising.

Prediction Markets: Kept His Distance, Mostly by Accident

Rabois did not invest in prediction markets. He attributes part of that to an early failed attempt at a company called Bluebat, where the core problem was not liquidity on the wager side but the inability to generate enough interesting bets at scale. He sees Polymarket's team as having played a difficult regulatory hand well, though he notes the company's historic challenge of US citizens using the platform before it was authorized. On the broader market category, he distinguishes between prediction markets with genuine social value — political markets, which he views as creating useful information liquidity — and platforms that simply expand retail gambling exposure for Americans who already have limited savings.