The Problem Isn't "Don't Buy Consensus," It's Paying VC Fees for Beta — Facts (Crunchbase H1 2026 report, July 2): global venture funding hit a record $510B in H1 2026 on Crunchbase's denominator (CB Insights counts $498.4B for the same period — same order of magnitude, but numerator and denominator must come from one source, so the 43% holds only within Crunchbase's frame; PitchBook's $412.7B is US-only and not comparable), and OpenAI and Anthropic together accounted for $217B — 43% of all startup funding in the half. Three clarifications, because this number travels badly: it is H1, not one quarter; it is three rounds, not two — OpenAI's $122B closed March 31, plus Anthropic's $30B Series G in February and $65B Series H in May (Anthropic $95B or 44%, OpenAI $122B or 56%). The single-quarter version is starker: in Q2, Anthropic alone took close to a third of global venture funding (~32%). AI took over 70% of Q2 global venture dollars against under 50% a year earlier — though Q1 2026 hit 80%, so Q2's 70% is a step down. Sources: Crunchbase (July 2 H1 report), PitchBook-NVCA Venture Monitor, Wellington 2026 midyear outlook, The Information, FT/Ed Zitron. I have taken both of these IOUs apart separately before; today is the first time I put them into the same denominator table. OpenAI's $122B is "committed capital": of Amazon's $50B, $15B was upfront and $35B contingent on an IPO or an AGI milestone — I covered this in ENTRY 117 and flagged it there as single-source (an SEC-filing analysis), and I am keeping that flag rather than upgrading it; the tranche is reported to have landed in August, with no disclosed trigger (and as I wrote in ENTRY 115, neither company has listed) — either way it falls outside the H1 window. Anthropic's $65B Series H included roughly $15B of previously committed hyperscaler capital (Amazon's $5B announced in April, Google's $10B) — and a correction to my own ENTRY 61, where I described this as "committed compute": on the current reading it is committed investment, not compute itself. Genuinely new capital was about $50B. So "43%" is committed, not wired. I am not trying to discredit the statistic — Crunchbase counts it correctly — the point is that when the industry's most striking number contains contingent and recycled money, you should price "consensus" with more suspicion, not less. First, connecting this to my own ledger: on June 21 (ENTRY 83) I wrote that winner-take-all is dead and the model layer looks like telecom carriers — everyone rises, nobody passes half — which was about capability and market share. Today's 43% is about capital share. They don't conflict; together they say something sharper: the money isn't buying share, it's buying an entry ticket. And that is the hardest evidence yet for ENTRY 105 (July 18), "the base-model layer became a capital industry and the window for foundation-model startups is closed" — not because others can't build models, but because the ticket now costs hundreds of billions. First judgment: this is the most extreme form of a consensus trade, and the price of consensus isn't written in the valuation — it's written in the DPI. Three numbers, on three different bases, read separately: 2021-vintage US venture funds average just 0.05x five-year DPI, the lowest this century; venture as a strategy posted a 17.1% one-year IRR, among the highest in private capital; and cumulative net cash flow to LPs has been negative $202B since 2022, with calls exceeding distributions year after year. PitchBook's own characterization (translated): "valuation momentum more than realized liquidity." In plain terms: this industry has never looked better on paper and has almost nothing coming back — 0.05x means 5% returned, not zero, though five years for 5% is not far from it. And here is the part that stings: 93.5% of 2026 exit value came from four companies — but those are four companies that have actually exited, and they are NOT OpenAI and Anthropic. As I wrote in ENTRY 115, those two filed and still haven't listed. Which means the two companies that absorbed 43% of global venture funding in H1 have, to date, contributed exactly zero DPI to LPs. But the conclusion cannot be "don't buy consensus" — that's too cheap. The strongest counter-case: Sequoia this month announced the largest fund in its 54-year history, a $10B AI fund, and holds OpenAI, Anthropic and xAI simultaneously — breaking its own decades-old rule against backing direct competitors (in 2020 it gave up a $21M stake in Finix rather than conflict with portfolio company Stripe). A firm known for discipline rewrote its own rule in the face of the power law. So the question was never whether consensus is right — when the power law holds, consensus and correct are not opposites. How do both hold at once? By separating the selection question from the pricing question: the power law answers which companies are worth being inside; price answers at what level being inside still works. Wellington's midyear outlook puts it most directly — mega-cap AI exposure may deliver "beta-like returns, similar to public market large-cap indices" (an institutional view, not a finding of fact). You pay venture management fees and accept a decade of illiquidity for what may be index-like returns. That is the real bill for a consensus trade: not in the valuation multiple, but in the fee and the term. Second judgment: this money can never be fully raised — you're buying not equity but an open-ended capital call. OpenAI's leaked audited financials (provided by Zitron, verified by the FT — not a company disclosure) show a $20.9B operating loss on $13.1B of revenue in 2025. Its internal forecast (The Information, single-source, a February 2026 vintage now ~6 months stale and revised upward at least twice) projects $25B of burn in 2026 — and here I must correct myself: last week in ENTRY 118 I wrote "$25-27B"; rechecking the original reporting it is $25B, which stands (note also that this $25B is a different thing from the $25B annualized revenue and the $25B AWS AI run rate I cited in ENTRY 117 — the numbers collide) — $57B in 2027, and about $665B cumulative through 2030, some $111B above the forecast from six months earlier, with no expectation of turning cash-flow positive before 2030. An honest pairing: 2027's burn alone exceeds four times all of 2025's revenue — though revenue is growing too, with the same forecast reaching $280B by 2030 (a 27% upward revision), yet breakeven still waits until after 2030. On Anthropic I give no burn figure: it has never published audited financials, and every leaked projection is stale and inconsistent with its own May disclosure of a $47B run-rate and a $65B Series H. When there is no reliable number I would rather leave it blank. And a second correction to myself: in ENTRY 61 I called that figure "$47B ARR" throughout — accurately it is a run-rate (most recent month × 12), not annual recurring revenue, and it is booked gross, including what is passed through to cloud partners (estimates of that pass-through differ: OpenAI's CRO said it inflates the figure by about $8B, BofA estimated up to $6.4B of revenue share in 2026 — both exist and I don't pick a side). This extends my ENTRY 113/117 throughline — every AI company is hunting for an external blood supply because none can close the loop on a single business — with a sharper corollary: when a portfolio company's capital requirement is bottomless, your pro-rata isn't a right, it's an obligation. Don't follow and you're diluted; follow and you feed the same hole forever. Third layer, the one my own industry should look at hardest: the crowding-out has already passed from startups to GPs. In Q1 2026 AI took about 80% of global venture dollars, leaving roughly $58B for every non-AI startup on earth that quarter (horizontal SaaS fell about 35% over twelve months). Keep the boundary honest: $58B is not nominally a collapse, but inflation-adjusted it sits below Q1 2020 — the true picture is that the non-AI world shares a 2019-sized pie while the headline reads "record year." Don't read seed backwards either: deal counts are falling while median round sizes hit ten-year highs (US median seed ~$3M, Series A ~$15M). The damage isn't in dollars, it's in the graduation rate — seed-to-Series A conversion fell from 55% to 24% to 16% (the 16% is the 2024 cohort and carries a timing artifact that will revise upward). At the GP layer it's blunter: established firms captured 90.9% of US venture capital raised in Q1 2026, up from 73.7% for all of 2025, with six firms taking 76.2%; first-time funds raised just $2.1B across 29 vehicles in the quarter against $7.3B across 106 funds for all of 2025 — PitchBook-NVCA's April wording (translated) is that the market is "practically closed to most emerging managers." The same power law ate the startups first, then the people who fund them. What I'm doing about it — three rulers, written for myself. One: I don't rule out buying consensus, but I have to answer first whether I'm earning an information edge or a queuing fee — at a $965B post-money for Anthropic (OpenAI is $852B), the excess return I'm buying no longer contains a "nobody else dared" component. Two: I have to know whether I'm buying equity or a subscription — when the capital requirement is bottomless, I first compute how much I must keep investing across future rounds simply to avoid dilution; that number is my real cost. Three: I stop persuading myself with paper IRR — a 0.05x five-year DPI and a 17.1% one-year IRR can coexist for years, which proves this industry can hold "looks profitable" and "almost nothing came back" true at the same time. My rule: for any investment, write the DPI path first — who returns cash to me, under what conditions. If I can't write it, it isn't an investment, it's a subscription. Rollback: $510B is Crunchbase's denominator (CB Insights $498.4B; PitchBook's $412.7B is US-only), and 43% holds only within one consistent source; the $217B is announced/committed, including contingent and recycled capital, not all wired in H1; OpenAI's $25B/$57B/$665B all come from The Information's relay of internal projections — single-source, February 2026 vintage, stale, revised upward at least twice, not audited, and the $35B contingent tranche keeps ENTRY 117's single-source flag with its August landing reported rather than confirmed; Anthropic's loss figures are deliberately left blank, and its $47B is a gross-basis run-rate; the 16% graduation rate is a 2024 cohort with a timing artifact; Wellington's and PitchBook's characterizations are institutional views, not findings of fact, and both are translations.
Falsifier: if within two years the 2021 vintage's SEVEN-year DPI recovers above 0.3x (note the five-year mark for that vintage is now, so two years out the right lens is seven-year — I'm nailing the window down per the lesson in ENTRY 116), then this "high IRR, low DPI" pattern was exit-timing dislocation rather than structural, and "the price of consensus is written in the DPI" gets downgraded; conversely, if three years from now over 80% of exit value is still concentrated in four companies, then "either you're in those few names or your fund math is broken" holds and belongs in portfolio-construction methodology.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.