20VC: Anthropic Surpasses OpenAI Revenue | OpenAI Acquisition of TBPN: Analysed | OpenAI Management Team Reboot | YC Kicks Delve Out | Mercor Hack and Why Now is the Time for Cyber | Supabase Raising at $10BN & Doug Leone Returns to Sequoia

The discussion centers on Anthropic reaching $30 billion in annualized revenue, up from $9 billion at the start of the year—a 3.3x increase in roughly four months. For context, Salesforce took 25 years to reach comparable scale; Anthropic has done it in approximately three to five years depending on the measurement baseline. Two structural advantages are highlighted. First, Anthropic remains compute-constrained, meaning demand exceeds supply—the company could sell more if it had more capacity. This has prompted a deliberate capacity allocation strategy: restricting high-token-consumption use cases on fixed-price plans (specifically, OpenClaw-type agent access) to redirect compute toward higher-revenue workloads. The broader principle articulated is that when a digital good is supply-constrained, rational economics dictates pricing closer to value while being careful not to suppress adoption, given that per-token costs will decline over time. Second, and more striking, Anthropic's model training costs are reported to be approximately one-quarter of OpenAI's. The discussion frames this as a compounding structural advantage: faster revenue growth, lower training costs, and a more focused product surface (no video, no consumer image generation) create a fact pattern that, in a public-market context, would likely attract long/short positioning—long Anthropic at its $370 billion valuation, short OpenAI at roughly $820–840 billion. The argument is that at roughly equivalent revenue trajectories, the valuation gap is difficult to justify given Anthropic's cost structure and growth rate.

OpenAI's simultaneous management overhaul—the CEO moved to "special projects," the CMO stepping down, the CRO departing, and the head of apps on leave—is read as a reactive response to competitive pressure rather than a proactive restructuring. The Oscar Wilde reference to losing "both parents" is invoked to describe the accumulation of departures, though the discussion is careful to separate health-related exits from strategic ones. The appointment of Denise Dresser, formerly CEO of Slack, to oversee go-to-market functions is characterized as a high-risk, low-base-rate move. The estimated probability of success when bringing in a "perfect LinkedIn" executive into a turbulent organization is put at roughly 30%. The core risk: there is insufficient time for a new leader to learn the product and organization before being handed a large portfolio during a period of competitive urgency. The acquisition of TBPN (a media/podcast network) draws the sharpest criticism. The argument against it is straightforward: OpenAI is arguably the most media-saturated company on the planet, with the CEO having access to every major world leader and dominating the AI news cycle. Buying a media asset with no editorial control provides no meaningful narrative leverage. The counterargument—that controlling media distribution matters strategically, as Andreessen Horowitz has argued—is dismissed on the grounds that editorial independence eliminates any practical benefit. The broader management lesson offered: the deal was likely conceived in January under different leadership priorities, and would almost certainly not be approved today. This illustrates a recurring M&A dynamic—management turnover kills deals in progress, and the window for any given transaction is narrower than it appears.

SpaceX has confidentially filed for an IPO targeting a $2 trillion valuation, which would surpass Saudi Aramco as the largest IPO in history. The reported financials: $15–16 billion in 2025 revenue, $8 billion in EBITDA, implying a revenue multiple of approximately 125x. The valuation incorporates the merged X/xAI entity, which is described as burning roughly $12 billion annually and holding at best a fourth-place position in the LLM market behind OpenAI, Anthropic, and Google Gemini. The discussion frames the gap between fundamental asset value and the $2 trillion target as "Elon Premium"—a premium that is visibly compressing on Tesla (down significantly year-to-date, with a major bank issuing a sell rating with a 60% downside price target). The IPO mechanics are expected to work as follows: a small float, 30% retail allocation, and the founder's force of will are likely to support the target valuation on day one. Whether it holds beyond that is treated as an open question. The broader structural observation is more significant for the venture industry: the combined IPO value of SpaceX, OpenAI, and Anthropic—assuming all three go public—is expected to exceed the total value of every other IPO over the past 20–25 years combined. This concentration at the top of the power law is described as psychologically disorienting for practitioners, though the practical advice is to avoid letting the scale of outliers redefine what constitutes a meaningful outcome.

Open Router—a marketplace that routes API calls across 50–60 LLMs dynamically, charging approximately 5% of the underlying model spend—has raised at a $1.3 billion valuation on $50 million ARR, up from $10 million ARR in October. The product is described as genuinely useful: it abstracts LLM selection complexity for application builders and enables dynamic model routing based on cost and capability. The valuation debate surfaces a broader concern about low-ACV (average contract value) AI infrastructure businesses. To generate $50 million in net revenue at a 5% take rate, Open Router is likely routing approximately $1 billion in inference spend. Scaling to $1 billion in revenue would require routing $20 billion in inference—achievable in theory given projected enterprise API spend, but dependent on capturing a large share of a market where major customers may eventually build direct integrations or shift to cheaper open-source models. Notably, the most popular models on Open Router are currently Chinese open-source models (Qwen, Kimi), meaning, as the discussion puts it, the Chinese Communist Party is effectively subsidizing American independent software vendors through subsidized open-source compute. The concern generalizes: AI investments with very low ACVs may have TAMs that are smaller in practice than they appear in projection, even as headline numbers look impressive. High-ACV AI companies (cited examples include Harvey and Lagora at $50–100K per contract) are currently flattering portfolio metrics in ways that low-ACV infrastructure plays cannot replicate. --- Key takeaways: - **Anthropic's training cost advantage (roughly one-quarter of OpenAI's) combined with faster revenue growth creates a compounding structural edge**—a rare combination where a challenger is simultaneously cheaper to operate and faster-growing than the incumbent. - **OpenAI's management reset carries meaningful execution risk**: bringing senior external executives into a turbulent organization during a competitive crisis has a historically low success rate, and the TBPN acquisition signals a focus problem that contradicts the company's stated code-red posture. - **SpaceX's IPO will likely achieve its target valuation on day one through retail demand and founder leverage**, but the inclusion of the xAI/X entity—a cash-burning, fourth-place LLM competitor—complicates the fundamental case for the $2 trillion figure. - **Compute constraint is now a genuine pricing signal**: both Anthropic and OpenAI are actively reallocating capacity away from high-token, low-revenue use cases, and this trend toward value-based token pricing will accelerate as demand continues to outpace infrastructure build-out. - **Low-ACV AI infrastructure businesses face a structural ceiling**: a small percentage take rate on a large but fragmented market may produce excellent products and strong early ARR without a clear path to venture-scale revenue, particularly if enterprise customers eventually consolidate or internalize routing decisions.