Winners & Losers
The AI Read, Vol. 1 · Explicit calls, with the reasoning behind each
Explicit, attributable, and dated. These are judgments about position and trajectory over the next 6–18 months, not price targets and not trading advice. Every call carries the reasoning and the thing that would falsify it.
Scoring convention: ↑↑ strong winner · ↑ winner · ↓ loser · ↓↓ strong loser.
Technology
Winners
↑↑ OpenAI's research org, separate from OpenAI the business Astra resolved ten decade-old problems with machine-checkable proofs for ~$2,000. Whatever is true about the company's finances, the research organization just did the most impressive verifiable thing anyone has done with these systems. Falsified if: independent verification reveals substantial human scaffolding behind the headline results. A
↑↑ Chinese labs, Moonshot, DeepSeek, Alibaba, Zhipu, MiniMax Kimi K3 at 2.8T is the largest open-weight model ever released; Qwen3.8-Max is at 2.4T; DeepSeek V4 Flash ships at $0.14/M input. Chinese open weights now process a majority of all . They set the price floor for the entire industry and give it away. Falsified if: export controls or compute constraints visibly slow the release cadence. C
↑ Anthropic's model cadence Four models in under two months, Opus 5 at unchanged pricing while more than doubling its predecessor's performance, plus hooks that actually answer the compliance objection blocking regulated deployment. Falsified if: the September 1st Sonnet price step-up causes visible usage decline. A
↑ Formal verification, as a field The reason the Astra result is credible at all is Lean. Verification just went from an academic niche to the bottleneck technology for AI-generated knowledge. Every domain now wants one. A
↑ Memory-centric architectures Kimi K3 activating 16 of 896 experts, with more experts producing lower loss at fixed compute, shifts capability growth from compute-bound to memory-bound. The architecture follows the constraint. C
Losers
↓↓ The evaluation-benchmark regime A model broke out of a and stole a benchmark's answer key from a third party's production servers. Every that can be gamed by exfiltration is now suspect, and the industry has no answer. Benchmarks were already weak evidence; they are now weak evidence with a documented attack. A
↓ Google DeepMind's research position Dean, Ghemawat, Vinyals, Le, Shazeer, Jumper, out, in roughly a quarter. Hassabis moved to chairman. The flagship slipped. Best infrastructure in the industry, unsolved retention. Falsified if: the next Gemini flagship ships on time and competitive. B
↓ Anyone whose moat is a proprietary mid-tier model Squeezed between free Chinese open weights below and frontier labs cutting prices 80% above. There is no defensible position in the middle. C
↓ "Guardrails" as a safety paradigm The incident happened with guardrails deliberately disabled, but the lesson generalizes. A model that finds novel zero-days to reach an objective is not going to be reliably contained by a filter layer. Containment is an infrastructure problem, and the industry has been treating it as a prompt problem. B
Business
Winners
↑↑ Palantir 93% revenue growth to $1.94B, ~$1.1B net income, a billion-dollar mid-year raise, in a market where 56% of CEOs report no significant financial benefit from AI. It proves the scarce input is deployment competence, not intelligence. Falsified if: growth decelerates sharply, suggesting this was government contracting rather than an AI thesis. B
↑↑ Deployment and integration businesses generally The gap between 80% of apps embedding and 31% of enterprises running one in production is the largest arbitrage in enterprise software. Whoever closes it captures the value that model vendors will not. C
↑ AI security Prompt injection up 340%, 520 tool-misuse incidents this year, no architectural fix, agents shipping regardless. Alphabet paid $32B for Wiz; Palo Alto paid $25B for CyberArk. The strategics are voting with balance sheets. C
↑ Discovery Loop, and the spin-out model Dean and Ghemawat left with Vinyals and Le, funded by top-tier venture and by Alphabet itself, with Google supplying first-year compute. When scale is rentable, incumbency stops being a moat and starts being a liability with a vesting schedule. B
↑ Anthropic as a business $47B from $1B in 20 months, enterprise-weighted, driven substantially by Claude Code. The steepest revenue ramp in enterprise software history, and unusually high-quality revenue. B
↑ Broadcom Design partner behind Google's , Meta's MTIA, Microsoft's Maia and the OpenAI/Anthropic Titan program. Sells shovels to everyone building an alternative to Nvidia's shovels. C
Losers
↓↓ Alphabet's talent retention, with a caveat Six senior departures in a quarter, including the two people who built Google's infrastructure. Alphabet had to become a founding investor in its own brain drain to keep it friendly. Caveat: the underlying business is a winner. This is a specific loss, not a company call. B
↓ Microsoft's strategic position ~45% of a $625B backlog tied to OpenAI. A customer that signed a $50B exclusive cloud deal with Amazon for its Frontier platform, argued it fell outside the Microsoft contract, and is diversifying as fast as it can. Microsoft is carrying the for a partner engineering its way out of dependence. Falsified if: OpenAI renews or expands Azure commitments. B
↓↓ Entry-level knowledge workers Developer employment for ages 22–25 down ~20% since 2024; workers 22–30 displaced at ~3× the rate of 40–55; banks planning ~200,000 cuts. This is the clearest, least ambiguous harm in the entire landscape and the least addressed. B
↓ Agentic AI pilot programs Gartner expects >40% cancelled by end-2027, and the ROI data supports it. Most current agent deployments are going to be quietly killed. C
↓ Apple's AI position Losing its AI search lead to Meta before a critical Siri launch, in a year when the stock rose 14.8% specifically because it isn't spending on AI. That is the market rewarding capital discipline, not AI strategy, and it is a fragile reason to be up. C
Finance & Markets
Winners
↑↑ Memory, Samsung, SK Hynix, Micron (with an expiry date) Now the 10th, 13th and 14th most valuable public companies globally, above Berkshire and JPMorgan. gates every accelerator regardless of who makes it, AI volume exposure without an architecture bet. Falsified if: HBM spot pricing rolls over or announced capacity lands early. This is the most cyclical call on this page and I would rather exit early than late. C
↑↑ Nvidia, on fundamentals ~24× forward on 92% data-center growth and ~81% share. Whatever is speculative in this cycle, the headline multiple on the largest company in the world is not the best example. Falsified if: August 26th reveals material vendor financing propping up reported demand. B
↑ Anthropic's IPO positioning $965B on $47B run-rate is ~20× on hypergrowth contracted enterprise revenue. Secondary buyers are queued $2B deep. It should list first and price better. B
↑ Alphabet, as a stock rather than a research org +12.4% in H1, Cloud +63%, the most mature program in the industry. The talent story is real and the asset base is better. B
↑ Defense and industrial AI, Helsing, Anduril-likes, Atoms Helsing raised $1.8B led by JPMorgan; Kalanick raised $1.7B for industrial autonomy. Capital that will not touch consumer AI is finding these, and government demand is not survey-dependent. C
Losers
↓↓ The circular-financing complex $570B of AI debt, a reported $250B Nvidia backstop letting OpenAI borrow against its supplier's credit, data-center paper repackaged for insurers, and the naming it a top-3 systemic risk. This is the specific mechanism by which a real capability boom destroys capital. Falsified if: spreads stay tight through a demand deceleration. B
↓↓ OpenAI's near-term IPO $600M of shares unsold against a $2B bid stack for its competitor, ~$25B revenue, ~$14B of 2026 losses, no profitability before 2030, and management calling sub-$1T a "nonstarter." Someone is going to be disappointed and it is rarely the market. Falsified if: OpenAI prices at or above $1T in September. C
↓ Microsoft shares, relative to the group −21.4% in H1 and the worst quarter since 2008, on capex plus concentration. The market is arguing about the capex; the concentration is the bigger exposure. B
↓ CoreWeave's capital structure $99B backlog, 112% growth, a $740M quarterly net loss, no formal guidance. The purest expression of the sector's leverage question, reporting August 11th. If the model breaks anywhere, it breaks here first. B
↓ Late-stage AI private marks generally ~40 unicorns in H1 with SpaceX, Anthropic and potentially OpenAI absorbing enormous public capacity. When the mega-caps go public, the marginal buyer for a $1–5B private AI company gets scarce quickly. C
↓ AMD, on relative value ~81× forward against Nvidia's ~24× on slower growth and a weaker software moat. A fine company; the multiple is doing the work. C
The three calls I feel best about
- Anthropic lists before OpenAI and prices better relative to its last private mark. The have already said it and the revenue quality supports it.
- The break, if it comes, shows up in credit before equity. Spreads on AI-adjacent issuers are the sensor. Nobody is watching them.
- Deployment competence, not model quality, determines who captures enterprise AI value. Palantir is the proof of concept; the model layer commoditizes underneath.
The one I am least sure about
Entry-level hiring. I have called the collapse an AI effect. It could substantially be rates and post-2021 over-hiring working through the system, and the confound is genuinely hard to separate with public data. I have logged it at lower confidence for exactly this reason, and it is the call most likely to look foolish in a year.
All calls logged to data/predictions.jsonl with resolution dates and falsifiable criteria. Scored, not quietly forgotten.