AI Finance Report
The AI Read, Vol. 1 · Markets, capital flows, IPOs, and where the opportunities sit
Research and analysis for a private readership. Not investment advice; no position recommendations are being made on your behalf. Figures are as-reported by the cited sources and marked by confidence tier, several are contested and shown as such.
1. The single most important development: the AI trade split in two
For three years "AI" was one trade. In 2026 it stopped being one.
The Magnificent Seven underperformed the index in H1, up 5.4% against the S&P's 7.9%, while dispersion inside the group went vertical: C
| Name | H1 2026 | Read |
|---|---|---|
| Apple | +14.8% | Rewarded for not spending. Briefly retook most-valuable-company on July 17th. |
| Alphabet | +12.4% | Best-positioned on infrastructure; then −5% on the Jeff Dean exit. |
| Microsoft | −21.4% | Capex plus OpenAI concentration. Worst quarter since Q4 2008. |
| Tesla | −11.3% | Optimus is real but internal-only; the AI narrative isn't carrying the stock. |
Meanwhile capital rotated within AI rather than out of it:
- Memory re-rated hard. Samsung, Micron and SK Hynix are now the 10th, 13th and 14th most valuable public companies globally, ahead of Berkshire Hathaway and JPMorgan. C
- Custom silicon inflected: reach 27.8% of AI server shipments in 2026, growing 44.6% against 16.1% for merchant . C
- Application layer produced the quarter's cleanest proof point in Palantir.
The generalizable lesson: the market has stopped paying for AI exposure and started paying for AI earnings. Microsoft and Meta were punished for spending without a legible revenue narrative in the same weeks Microsoft was rewarded (+16% in the July 28th–31st earnings week) for connecting spend to validated demand. Same companies, same , different story discipline. C?
Caution on the Microsoft numbers: the −21.4% H1 figure and the +16% earnings-week move are both reported and not contradictory, but they describe different windows. Sources also give Microsoft's YTD as roughly −12% at other points. Treat the direction as reliable and the precise level as. ?
3. Nvidia: the load-bearing print
| Metric | Value |
|---|---|
| Market cap | ~$5.31T, world's most valuable company |
| Milestones | First to $5T (Oct 2025); $5.5T (May 2026); briefly passed by Apple (17 Jul 2026) |
| Q1 FY27 revenue | $81.6B |
| Data center | $75.2B, +92% YoY, now 92% of total revenue |
| Q2 FY27 | August 26th 2026, ~$91B |
| Share | ~81% of data-center chip revenue; ~70% of AI chips overall |
| Forward P/E | ~24× |
A 24× forward multiple on 92% growth is not a bubble multiple. Whatever is speculative in this cycle, Nvidia's headline is not the clearest example, AMD trades near 81× forward on far slower growth.
The actual risks are downstream of the multiple:
- Customer concentration. A handful of buyers, several of whom are also building competing silicon with Broadcom.
- Financing quality. If Nvidia is backstopping customer purchases, revenue quality degrades even as revenue grows. This is the single most important thing to watch on August 26th, and it will be in the disclosures rather than the headline number.
- Memory gating. constraints cap unit shipments regardless of demand.
4. The IPO pipeline
Completed
SpaceX, June 12th 2026. 555.6M shares at $135. Surged on open, clearing $2T. The proof that mega-cap private companies can access public markets at scale, and the event that opened the window for everything below. C
Filed
Anthropic: confidential , June 1st 2026.
| Metric | Value |
|---|---|
| Valuation | $965B post-money (set by $65B Series H) |
| Revenue | $47B (late May 2026) |
| Trajectory | $1B (Dec 2024) → $9B (end-2025) → $47B: 47× in 17 months |
| Implied multiple | ~20× revenue |
| Possible listing | As early as October 2026 |
| Driver | Enterprise adoption and Claude Code |
Contested. Outlets have variously reported $380B, $965B and $1.75T for Anthropic. The $1.75T figure appears to be contamination from SpaceX's valuation in at least one report. $965B paired with $47B is the internally consistent, best-sourced pair, and it is the one used here. ?
OpenAI: confidential S-1, ~May 22nd 2026.
| Metric | Value |
|---|---|
| Target valuation | $730–850B; Altman and CFO Sarah Friar call sub-$1T a "nonstarter" |
| Original target date | September 2026 |
| Current status | Reportedly leaning toward delaying to 2027 (as of June 25th) |
| Revenue | ~$25B annualized (Feb 2026), from $6B in 2024 |
| Losses | ~$14B in 2026; profitability not expected before ~2030 |
| Banks | Goldman Sachs, Morgan Stanley |
The secondary market has already voted
$600M of OpenAI shares sitting unsold. $2B of bids queued for Anthropic. C
This is the most information-dense datapoint in the issue. Secondaries are where informed holders (employees, early investors, people who have seen the numbers) express views they will not state publicly. Unsold inventory on one side and a two-billion-dollar bid stack on the other, at a moment when both companies are pre-IPO, is a verdict.
Why the divergence is rational: Anthropic shows ~$47B run-rate at 20× with enterprise- weighted, contracted revenue. OpenAI shows ~$25B at a proposed ~34× with consumer-weighted revenue, $14B of annual losses, a customer relationship with Microsoft that it is actively restructuring, and no profitability path before 2030. On quality of revenue: it is not close.
5. Private markets
- ~40 AI unicorns created in H1 2026, valuations $1B–$41B. C
- Prometheus (co-founded by Jeff Bezos): $12B Series B, the largest single round of the period; $18.2B total, $41B valuation. C
- Databricks: raising at $188B, led by Coatue (July 16th). B
- Helsing: $1.8B (July 13th), led by JPMorgan, Lightspeed and Iconiq. Defense AI continues to attract capital that will not touch consumer AI. B
- Atoms (Travis Kalanick), $1.7B for industrial AI. C
- Figure AI: $39B valuation, $1B+ raised, 10,000+ deployments. C
- Cursor/Anysphere: crossed $2B annualized revenue in March 2026. C
- PsiBot (China), ~$100M at $1.48B. Small, but a marker for Chinese robotics capital formation. B
M&A: 266 AI deals in Q1 2026, +90% YoY. Nearly half of all strategic tech deal value above $500M now comes from AI-native targets, up from ~25% in 2024. Alphabet/Wiz ($32B) and Palo Alto/CyberArk ($25B) are the large prints, note that both are security. BC
6. Catalyst calendar
| Date | Event | Why it matters |
|---|---|---|
| 11 Aug | CoreWeave Q2 | Est. rev $2.56B vs $1.21B YoY; $99B backlog against a $740M quarterly net loss and no formal . The purest leverage stress test in the complex. |
| 26 Aug | Nvidia Q2 FY27 | Consensus ~$91B. Watch financing disclosures, not the headline. |
| 31 Aug | Claude Sonnet 5 promo pricing ends | $2/$10 → $3/$15 on 1 Sept. First real test of frontier pricing power. |
| Sept | OpenAI IPO window | If it slips, the secondary signal is confirmed. |
| Oct | Possible Anthropic listing | Would be among the largest IPOs ever. |
| Ongoing | White House frontier framework | Determines whether "Pacing the Frontier" becomes statute. |
| Ongoing | Astra proof verification | Independent verification confirms or punctures the year's biggest claim. |
7. Where the opportunities and risks actually sit
Framed as theses with explicit disconfirming evidence, so they can be scored later.
Thesis 1. Memory is the least-crowded bottleneck (high conviction)
Every accelerator needs HBM; HBM is capacity-constrained; the constraint gates Nvidia and its competitors equally, which makes it a bet on AI volume without a bet on which architecture wins. Microsoft attributing $25B of capex to memory inflation is the pricing power showing up in someone else's income statement. Against it: memory is the most violently cyclical business in semiconductors, these names have already re-rated substantially, and capacity additions are being announced now. Disconfirmed by: HBM spot pricing rolling over, or credible capacity coming online ahead of schedule. C
Thesis 2. Custom silicon is a structural share shift, not a cycle (high conviction)
ASICs at 27.8% of shipments growing 44.6% against 16.1% for merchant GPUs is not noise. Every hyperscaler is simultaneously Nvidia's biggest customer and its most motivated competitor. Broadcom sits behind Google , Meta MTIA, Microsoft Maia, and the OpenAI/Anthropic "Titan" program. Exposure to the migration without picking a winner. Against it: frontier training stays CUDA for the foreseeable future; the migration is inference-led and slower than shipment share implies. Disconfirmed by: a major hyperscaler publicly deprecating its internal program. C
Thesis 3. Credit, not equity, is where this breaks (high conviction on mechanism, low on timing)
$570B of AI debt, utilities levering into data-center demand, vendor backstops letting weaker credits borrow against stronger ones, and data-center paper being repackaged for insurers. Equity markets are reasonably priced on earnings; the leverage is in structures most equity investors are not reading. Watch: credit spreads on AI-adjacent issuers, utility issuance, and any disclosure of vendor financing in Nvidia's August 26th filing. Against it: the underlying assets generate real cash flow, unlike 2001 fiber or 2007 mortgages. Demand is not speculative: it is contracted. Disconfirmed by: spreads staying tight through a demand deceleration. B
Thesis 4. Enterprise AI winners are deployment companies, not model companies (medium-high)
Palantir at +93% revenue growth and ~$1.1B net income, in the same market where only 23% of enterprises report significant ROI and Gartner expects 40% of projects cancelled, says the scarce input is not intelligence: it is the ability to make it work inside a real organization. Models commoditize toward the Chinese floor; deployment does not. Against it: Palantir's multiple already prices a great deal of this, and its government concentration is a different business from commercial enterprise AI. C
Thesis 5. Microsoft's real risk is concentration, not capex (medium)
~45% of a ~$625B backlog tied to OpenAI. A single customer that signed a $50B exclusive cloud deal with Amazon for its Frontier agent platform, arguing it fell outside the Microsoft contract. Microsoft's stake was restructured to ~27%. The market is pricing the capex; the concentration is the larger exposure. Against it: the backlog is contracted and OpenAI's compute needs are growing faster than it can diversify. Disconfirmed by: OpenAI renewing or expanding Azure commitments at the next milestone. B
Risk that is under-priced: the security event
Prompt injection up 340%, 520 tool-misuse and privilege-escalation incidents in 2026, and a documented case of a autonomously breaching a third party's production infrastructure. Agent deployment is far outpacing agent security, with no robust architectural fix available. Alphabet paying $32B for Wiz and Palo Alto $25B for CyberArk suggests strategics see it clearly even if the market doesn't. C
What I would not touch
Pure-play "AI" companies whose moat is a model they do not own, competing against free Chinese open weights at a tenth of Western pricing. That floor is falling, and an 80% price cut on GPT-5.6 Luna is the incumbents telling you where it goes. C
8. The frame I'd hold
A capability boom and a capital bust are not mutually exclusive. Telecoms in 2001 is the correct analogy and it is usually deployed lazily: the fiber was real, the demand forecasts were eventually correct, and the equity was still destroyed, because the build-out was debt-financed against demand that arrived years after the interest payments were due.
The AI build-out is better than that in one crucial respect: the demand exists today and is contracted. Nvidia's data center revenue is not a forecast. It is worse than that in another: the financing has become circular in ways 2001's was not. When a chip vendor backstops a customer's ability to lease capacity to run the vendor's chips, revenue quality and credit quality are the same variable wearing two hats.
Both the boom and the fragility are real. Position for the technology being right and the financing being wrong. Those are separable bets, and most of the market is treating them as one.