The Investor

Two incompatible explanations, one identical consequence

Altman's stated reason is alignment and safety work; he explicitly denied capital pressure. His own CFO told employees in August that the company "will be a public company in 2027," and a confidential S-1 went in during June. The New York Times separately reported that bankers advised against listing, citing SpaceX-style volatility and unclear finances. Pick either account and the structural result is the same: no date, no banker, no structure, and a gating condition with no accepted standard to clear.

What that removes is not liquidity — employee tenders and ROFR-gated secondaries still function. It removes price discovery. The listing was the only event capable of putting a public number on frontier AI. Without it, private marks stay model-derived and unfalsifiable for another 12-24 months. Two consequences follow, in order of how fast they hit a book. IRR degrades on timing alone: TVPI holds, the clock does not, and any position underwritten to a 2026-2027 frontier comp just lost 18-24 months of return velocity. And repricing, when it comes, arrives through structure rather than tape — down rounds, layered preference, secondary discounts. Slower, quieter, and with no floor under it.

Who sets the reference now

Three parties, none obliged to publish a methodology: secondary buyers, tender agents, and lenders holding private shares as collateral. SoftBank is the cleanest example of the last. It has committed $64.6B for about 13% of OpenAI, financed partly with a $10B margin loan and an $11.87B two-year facility, has repaid a $40B bridge down to $25.9B, and has a final $10B tranche due October 1. Its preferred shares convert to common only at an IPO, so the deferral postpones the conversion event embedded in the financing.

Here is the mechanic most books miss. A margin loan against private shares has no daily price. Lenders fall back on the last primary round, secondary indications, or a negotiated haircut — so a change in OpenAI's reported valuation can hit a balance sheet before a single share trades. SoftBank is marketing a possible $10-20B high-yield deal in New York.

That bond print is the market's stated opinion on how to discount an undated private AI stake. It is public, it is dated, and almost nobody will use it as a discount rate.

Where the reporting diverges

Coverage agrees the print is gone and that secondaries become the clearing mechanism through 2027. It splits on whether the issuance window itself is shut, and there is counter-evidence: Electra Therapeutics pricing up to $347.2M on Nasdaq at $14-16 and Bamboo Insurance up to $700M on NYSE at $18-20, backed by OrbiMed and by CVC and White Mountains respectively. Markets are open. What is closed is access for issuers carrying frontier-AI narrative risk — an asset-specific disclosure problem, not a market-wide freeze.

So stop treating the last private round as a reference and start deriving one. The bond market will price undated private AI collateral in that offering; a distressed strategic buyer priced horizontal software. Both are harder numbers than any up-round sitting in a September mark.

Run Oracle's arithmetic before you accept the narrative

Oracle removed roughly 21,000 people — about 13% of its workforce — in the fiscal year ended May 31, and revised its 2026 restructuring plan up $700M to $2.8B, largely severance. Even at $200K fully loaded, 21,000 heads is about $4.2B a year of payroll relief against $90-95B of planned capex — under 5% of the spend. The layoffs are not financing the buildout; $125B of debt is, carried at a lower credit rating than rival hyperscalers, with the headcount cut serving as a margin signal aimed at bondholders.

The read-through for a private book is direct. If an issuer with an enterprise installed base and a contracted backlog is stretching this far, every unrated neocloud, GPU lessor and data-center developer in the pipeline is financing on worse terms than its model assumes. Credit capacity, not chip allocation, is the binding constraint at the second tier.

Then run the physical one

Moody's quantified the other half: $110B for 45 gigawatts of incremental US generation through 2030 to serve data-center load, roughly $2.44B per gigawatt, per Bloomberg. Dispatchable supply is shrinking while load grows — 330 coal plants retired since 2010, with 60 more slated by 2031.

The permitting leg is where consensus is furthest behind. Gallup found 7 in 10 Americans oppose a data center near them, over half citing environmental harm, and indebted rural towns are now refusing substantial promised tax revenue outright. The federal response has been to frame that opposition as a foreign influence operation rather than accommodate it, which does nothing for a permitting calendar. Every infrastructure asset needs three numbers rather than one — energized, contracted and permitted megawatts — plus an IRR re-run with a twelve-month interconnection delay. The spread between a developer with a signed interconnect and one with land plus a press release is about to widen sharply.

The policy trade cuts both ways

The EPA rolled back carbon standards for fossil-fired power plants with a claimed $300B+ of industry savings, eliminating the coal carbon-capture-by-2039 requirement, and proposed that greenhouse gases do not endanger human health — an attack on the scientific predicate itself, which is a bid for durability across administrations. Two opposite consequences follow. Any climate position whose demand driver is a mandate rather than economics needs re-underwriting at zero policy support. And dispatchable, behind-the-meter generation gains margin exactly as data-center load strains grids. Environmental groups are expected to litigate, so the $300B is contingent and the timing uncertain — underwrite unsubsidized cash flow, not the rule.

Where the readings disagree

Three independent analyses converge on the scarce asset — a permitted, energized megawatt and the credit to build it — and split on who captures the value. One holds it accrues to owners of energized, permitted capacity. Another notes that power still trades at infrastructure multiples rather than AI multiples, which is the actual arbitrage. The third points somewhere adjacent: Marsh expects AI campuses to outgrow conventional insurance capacity, in a captive market already running 6,000+ captives and roughly $240B in premiums. That is compute-adjacent growth available at depressed insurtech entry pricing, and the least crowded idea on the board today.

Where the demand is legislated and unpriced

The most underwriteable position in this material is not a missile. The SASC markup — advanced 18-9 on June 11 — requires second sources for solid rocket motors and directs accelerated acquisition across eight critical categories. Solid rocket motors, seekers and energetics are the named production bottlenecks that money cannot compress. Monopoly incumbents, a statutory buyer, and no valuation premium attached yet.

The number that governs every mark in the sector

The FY27 NDAA that cleared the Senate Armed Services Committee in June authorizes $1.15 trillion — the largest topline in history — with a requested 188% increase in missile procurement. Against that, the Munitions Acceleration Council's priority list runs above $47B, and less than 3% of the munitions plan funds low-cost munitions. Do the arithmetic the advocacy leaves implicit: the entire funded new-entrant opportunity today is roughly $1.4B. Every deck anchors on the topline. None anchor on the 3%.

Offensive strike is consolidating, not forming

More than half a dozen competitors now sit in the $300K-$700K band with contracts signed and deliveries scheduled. The Air Force's FAMM program sets the anchor at 28,000 weapons for $12.6B — about $450,000 average unit price, the figure the government will negotiate everyone toward — and the Navy's MACE program carries a hard $300K ceiling. A crowded supply side facing a government-set price ceiling and a monopsony buyer is a price war, not a formation-stage market.

The counterintuitive winner

Read the production tables rather than the polemic: prime munitions revenue is inflecting violently upward. PAC-3 MSE goes from roughly 600 to 2,000 units a year at $3.9M — about $7.8B of annual run-rate. THAAD quadruples from 96 to 400 a year at $12-22M. SM-6 goes from 125 to 500-plus. The interceptor scale-up alone is a larger incremental revenue pool than the entire low-cost strike segment, and it accrues to Lockheed and Raytheon. Displacement is real but slow: rebuilding major-program inventories takes 8-14 years at historical rates, complex missiles take two years to build, and plant expansions take three.

Revenue quality, not revenue, is the diligence axis. One study found munitions funding swinging more than 50% year-over-year across eighty programs. When the Army bought only 23 JAGM missiles in FY25, unit price hit $360,000 against a $212,000 program-life average. New entrants financing production off their own balance sheets absorb that volatility as inventory, idle capacity and working-capital burn.

A framework agreement without obligated quantities is not a backlog. Discount it at least 70% and price the funded programs of record separately.

Two discounts to apply before you believe any of it

Base rate first: the Collaborative Combat Aircraft was supposed to cost $3M and now costs $25-30M. Haircut any pre-delivery unit-price claim 2-3x, and remember that a $200K weapon with half the reliability is really a $400K solution. Second, the conflict is disclosed and material: the three companies presented as the answer are all a16z portfolio holdings, and the "1/5 to 1/10 of a Tomahawk" cost figure is author-sourced. Treat the policy analysis as excellent and the cost claims as marketing until the comptroller's weapons-price exhibit confirms them.

The sector bid is genuine and pact-proof: Shield AI reportedly moved from $12.7B in March to $20B+ in talks, Ursa Major sits at $2.3B, and six defense and space SPACs have priced year to date against three in all of 2025 — roughly 10% of all SPAC deals. Budget-driven demand does not care about model-release cadence. It also does not care about your fund's hold period, which is shorter than the 7-10 years this converts over.