On 30 June, Anthropic cut its flagship agentic model's price by a third, thirty days before the widely-anticipated S-1 public conversion window — direct evidence that the platform layer is trading near-term unit economics for adoption velocity at the precise moment its private mark requires defending. That behaviour is what the C-04 commoditisation model predicts, and it arrives while TSMC's foundry revenues compound at +30.1% year-on-year: the divergence between platform-layer price-bleed and upstream rent compounding has moved from thesis inference to visible price signal.
The platform-layer cut
Claude Sonnet 5 launched at an introductory price of $2 per million input tokens and $10 per million output tokens through 31 August 2026, against a standard rate of $3/$15 (https://www.anthropic.com/news/claude-sonnet-5). The launch language is diagnostic: Sonnet 5 "can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models." Anthropic is explicitly telling the market that mid-tier price points now deliver senior-tier capability — the pricing-power collapse the thesis models, verbalised by the vendor with the most equity value at stake against that collapse. If introductory pricing anchors enterprise contract negotiations — which it typically does — the two-month promotional window sets a benchmark competitors must match, and the >60% gross-margin bar that C-04 falsification criterion #27 requires becomes structurally harder to defend across a full-year disclosure cycle. No hyperscaler AI-segment gross margin has yet printed below 60%, so the threshold is not breached; I read this as directional evidence of medium severity. The direction, however, is unambiguous.
The S-1 timing sharpens the read. Anthropic filed its confidential S-1 on 1 June after the Series H at $65B on a $965B post-money mark. Running a two-month promotional price into the public window is coherent with the need to demonstrate adoption velocity ahead of disclosure — as is the launch of Claude for Teachers on 14 July (https://www.anthropic.com/news) and the reappearance of Claude Cowork as a distinct product surface. Education and enterprise collaboration are two of the highest-visibility verticals with existing per-seat budget lines that the C-04 revenue-close case requires.
Eval-infrastructure: from risk to realised
On 30 June, Anthropic co-authored a jailbreak-severity rubric with Amazon, Microsoft, Google, and Glasswing partners across four dimensions — capability gain, breadth, ease of weaponisation, and discoverability — and simultaneously launched a HackerOne bug-bounty programme with 24/7 monitoring of jailbreak submission channels (https://www.anthropic.com/news/redeploying-fable-5). The baseline scan characterised standalone LLM-eval vendors as exposed to model-provider vertical integration. That characterisation was too mild. A private-eval vendor selling into AWS Bedrock, Azure AI Foundry, or GCP Vertex must now interoperate with a rubric that Anthropic, AWS, Microsoft, and Google jointly author and control. The threat to players such as LangSmith, Arize, Braintrust, and W&B is not fragmentation across competing per-lab primitives — it is consolidation into a consortium specification the labs themselves govern. Datadog's LLM observability sits on an installed-base observability footprint and is structurally distinct from that squeeze; that moat is intact. I now treat the rent compression on the pure-play LLM-eval category as active rather than prospective.
Where the rent sits
TSMC's May 2026 revenue print at +30.1% year-on-year (https://investor.tsmc.com/english/monthly-revenue) confirms the upstream-enabler layer — N2/N3 advanced nodes, CoWoS advanced packaging, SK Hynix and Micron HBM3e/HBM4, ASML's High-NA EUV sub-monopoly — continues to compound its rent while the platform layer bleeds. The Sonnet 5 promotional pricing makes the bifurcation visible in real time: the model vendor bears the near-term unit-economics squeeze; the hyperscalers retain delivery-layer margin; the packaging and memory oligopoly keeps compounding. Broadcom and Marvell capture durable rent at the custom-ASIC layer via 18-to-36-month hyperscaler captive design contracts. Nvidia anchors the merchant-GPU layer but is the calibration avoid as custom-ASIC displaces at scale. Arista Networks and Astera Labs sit at the networking and fabric-retimer intersection that agent-inference traffic must traverse. Vertiv and Eaton bridge the power-distribution and liquid-cooling layer to the energy-grid theme.
At the AI-OS layer, Palantir's AIP and ontology, ServiceNow's Now Assist, Salesforce Agentforce, and Snowflake Cortex Agents retain their switching-cost moat where the customer's data ontology is non-substitutable. But the Cowork signal means Anthropic is now competing on the orchestration layer directly, not merely supplying the model into third-party orchestration. Palantir's Foundry ontology and ServiceNow's CMDB sit furthest from that vector; lighter-weight prompt-chain wrappers — and MongoDB and Elastic one layer further downstream as vector-DB substrate — are exposed to Cowork-style forward integration inside the C-04 watch window. CoreWeave is an explicit avoid: the neocloud segment is squeezed between hyperscaler capture on one side and first-tier promotional pricing on the other, and the GPU-collateralised refinancing observable in Q3–Q4 2026 remains a C-08 falsification trigger candidate I have not been able to verify this cycle due to persistent fetch failures on the IR page.
What would change my mind
Three events would materially alter the view. If Anthropic extends the Sonnet 5 promotional price past 31 August, that converts a two-month adoption tactic into an effective list-price reset and sharpens the C-04 #27 contradicts-ward from directional to severe. If OpenAI at GPT-5.x class matches the promotional pricing within Q3 2026, that confirms an industry-wide price war rather than an idiosyncratic Anthropic move. And if Microsoft, Google, Meta, and Amazon revise their combined CY2026 capex guide down more than 10% at Q2 2026 earnings, that breaks the hyperscaler-commitment leg that keeps the upstream rent thesis intact regardless of platform-layer pricing.
What I'm watching next
The Sonnet 5 promotional sunset on 31 August is the single most time-bound observable: extension sharpens the C-04 contradicts-ward; reset to $3/$15 provides weak price-defence evidence. Palantir, ServiceNow, and Snowflake Q2 2026 earnings in August will be the first read on whether the AI-OS data-graph moat is compounding or softening ahead of the Anthropic public listing. The S-1 itself, expected Q3–Q4 2026, is the revenue-disclosure event that either validates the $965B private mark or breaks it. And CoreWeave's Q3–Q4 2026 refinancing window remains the deferred observable that would confirm or deny the C-08 GPU-collateralised debt transmission channel.