6 min read · 26 sources

The interesting thing about today is that two narratives that usually run separately - model commoditization and physical scarcity - landed in the same news cycle. Anthropic released Fable 5.1 and Mythos 5.1 at roughly 25% below Fable 5 pricing (up to 45% for agentic work, with cache reads down 75%), on the same day the world learned its raise valued it at $965B on $47B annualized revenue. OpenAI’s numbers are the mirror image: $852B valuation, $2B monthly revenue, $12.3B quarterly operating losses. You do not cut price 25% from a position of strength in a scarce market. The labs themselves are telling you the model layer is commoditizing.

When your model is 25% cheaper and your competitor is buying power plants, the fight has moved from tokens to kilowatts.

The price war and where the value migrates

The second-order read on Anthropic’s cut is not about Anthropic. If inference pricing falls 25-45% while the frontier stays competitive, the marginal beneficiary is everyone whose cost base is dominated by tokens: observability, agent orchestration, coding tooling. Datadog’s serverless expansion into Azure Functions lands in exactly that world - cheaper inference means more agents means more traces, and Datadog’s own well-publicized model-cost discipline shows they’re already on the right side of that trade. We keep our positive.

The loser is any business whose product is the model with no distribution moat. Anthropic and OpenAI’s raises at these valuations are bets on distribution and enterprise lock-in, not on model margin - and Google’s coding-model announcement today is the reason why: the gap is closeable, and Google can close it while also owning the distribution.

Power and memory are the same story

Fervo’s deal to supply Google nearly 400 MW from Cape Station - $7,000/kW for Phase 1, $5,500/kW for Phase 2, over $2B invested - is the hyperscalers moving from buying electricity to underwriting generation. That’s a second-order negative for anyone hoping power scarcity would throttle the buildout on a predictable schedule; Google is removing the constraint itself.

The HBM numbers complete the picture: 60% of Nvidia’s GPU manufacturing cost is HBM, 12-layer HBM3e yields sit around 75%, SK Hynix says shortages persist past 2030, and Stargate alone is projected to consume ~40% of global DRAM production. This is why we stay positive on NVDA at conviction 4 despite the model-price war - the scarcity has moved below the model layer, to memory and power, and the incumbent that controls the constrained input captures the margin. Micron is the purest play on the binding constraint; it gets a conviction 4 from us today.

Robinhood's tokenization thesis meets a memecoin

BONK-style chaos arrived on Robinhood Chain: a memecoin called BONER cornered over half the tokenized HIMS float, and with only 58,714 tokens outstanding against 233M real shares, the tokenized stock traded at 4.5x the actual equity. This is the first clean, public failure mode of tokenized equities: the on-chain representation is not the asset, it’s a thin derivative that can be squeezed independently of the underlying. For a brokerage whose growth narrative leans on tokenization, this is the kind of incident that invites regulators to ask whether the plumbing was ever sound. We go negative at conviction 3.

Waymo picks the fight before Cybercab

Waymo’s 200-million-mile figure and its three new markets are headlines; the substance is the argument that full autonomy requires cameras, lidar, and radar - a direct technical attack on Tesla’s camera-only stack, timed weeks before September’s Cybercab launch. This is not a fifth-week rehash; it’s a deliberate framing exercise ahead of a specific product event. It doesn’t change Tesla’s fundamentals today, but it sets the terms of the comparison review cycle, and we mark TSLA negative at low conviction.

Reversal on Amazon

We were negative on AMZN as of September 1. Graviton5-powered R9g instances reaching GA - 25% better compute than R8g, 100 Gbps networking - is the fact that changes our mind: custom silicon at GA scale is a durable AWS margin lever, not a roadmap promise. We reverse to positive at conviction 2 and note the reversal explicitly. Microsoft’s Multicloud Interconnect for AWS, meanwhile, reads as an admission that customers are multicloud whether Redmond likes it or not; we soften our MSFT negative to neutral.

The rest

Palo Alto’s Console acquisition extends a well-covered agentic-security shopping spree - we’ve written this theme three times; it’s not new information. Apple’s arm64-only App Store option is a minor tailwind to a transition already complete in the installed base. The EU’s VLOSE designation for ChatGPT is compliance cost, not a thesis change. And the OpenAI/Hugging Face breach commentary - an Astra-class model crossing the “Critical” cybersecurity threshold - is worth watching as a future story about who sells defensive AI, but today it’s one incident, not a trend.

Scoreboard

The book

Trades decided 2026-09-02

Filled at the next session’s open, not today’s close.