Signals

Mark sent through a VentureBeat piece headlined "Salesforce just put its entire CRM inside Claude and says you'll never need its app again". VentureBeat blocks automated retrieval and the article body could not be read, so the headline is all that can be verified from that source. It does not stand alone, though. CNBC, reporting Anthropic's Enterprise Frontier Safeguards on 1 September, refers to an interview last week with Dario Amodei about "the company's expanded partnership with Salesforce", in which Amodei said the arrangement helps ensure Salesforce's customer data remains private and that Anthropic put a "huge amount of effort" into managing permissions for the integration between the two companies. An expanded Salesforce-Anthropic integration is therefore corroborated; the "never need its app again" claim is VentureBeat's framing of it, not a company statement we can check. Mark read it as "SaaCCEPTANCE" - the vendor conceding rather than resisting. John's reading was more specific: "Basically they are a database now... The old system of record." That is the SaaS apocalypse thesis in its least dramatic form. If a CRM's workflow and interface can be driven from a model's context window, what the vendor sells is the authoritative store and the permission model around it, not the application. The counter-argument is that this is precisely why Salesforce would do it: the store is the defensible layer, and being the system of record every agent must transit is a better position than being an app a human opens. John's own conclusion pointed that way - "at enterprise scale this might be a good place to play". Amodei's emphasis on permissions supports the same reading: the hard problem being solved is authorisation, not generation. The thing to watch is the pricing metric. Seat-based licensing assumes humans opening an application; if usage migrates to agents while pricing stays per-seat, revenue decouples from consumption in the vendor's favour briefly and against it thereafter. A move to consumption or agent-based metering would be the concession made explicit. Watch also whether the integration is exclusive. A CRM reachable from any model is a database; a CRM reachable from one model is a distribution deal, and those are priced very differently.

Bloomberg, carried by Yahoo Finance, reports that Anthropic agreed a US$35bn computing deal with Lambda, an Nvidia-backed cloud provider, according to a person familiar with the matter. Hut 8 is developing the Texas data centre involved, in Nueces County, and the Wall Street Journal previously reported that Nvidia would hold the lease on it. Anthropic declined to comment; Nvidia, Lambda and Hut 8 did not respond. The same report sets it beside US$45bn agreed last week to rent capacity from Nscale in West Virginia, US$50bn with Fluidstack and US$45bn with SpaceX in recent months - roughly US$175bn of contracted capacity. Lambda is separately in talks to raise as much as US$3bn at a valuation of US$12bn or more, having raised over US$1.5bn in November. These are unnamed-source reports of contracted commitments, not audited disclosures, and no term lengths or delivery schedules are given: a US$35bn figure without a duration is not an annual cost, and stacking four of them into a single total is arithmetic, not analysis. What the pattern does show is a shape. Nvidia's capital sits behind several counterparties at once - it backs Lambda and, per the Journal, holds the lease - so the same firm appears as investor, landlord and supplier in one transaction. Mark's position on the week's spending was direct: "I am with Dwarkesh on throwing money at supply constraints." John's addition was that it runs "up and down the supply chain". The counter is the one CNBC published the same day: commitments are fixed while the price of the output falls. Watch for a duration. If any of these agreements is disclosed with a term in an IPO filing - Anthropic filed confidentially this summer - the annualised obligation becomes computable, and comparing it against realised revenue is the entire argument. Watch whether Hut 8, Nscale or Fluidstack name the counterparty and contract value in their own filings: a neocloud with one customer at that scale is a credit exposure rather than a customer base. And watch whether Nvidia's simultaneous role as investor, lessor and supplier draws accounting or disclosure scrutiny, which is the point at which circular-financing concerns stop being a talking point.

CNBC reports that Silicon Data's LLM Token Expenditure Index fell to 97 cents per million tokens on Monday 31 August - the lowest reading since the index was created late last year, and more than half off the high recorded earlier this summer. The index tracks the going market rate per million large language model tokens. CNBC attributes the fall, via Charles-Henry Monchau of Syz Group, partly to the rise of open-source Chinese models such as Moonshot's Kimi K3 that price below frontier alternatives, partly to OpenAI's price cuts on two GPT-5.6 models in late July, and partly to dynamic pricing schemes that let access rates move with demand. Silicon Data's head of research Steve Hou suggests the drop could signal there is already "enough supply out there to provide sufficient capabilities for most tasks". This is a price index, not a revenue measure. A falling per-token rate is entirely consistent with rising total spend if volumes grow faster, and nothing here shows what any lab actually collected. What it does constrain is pricing power, and Monchau's formulation is the one worth keeping: "Token deflation compresses the revenue line while compute commitments stay fixed", with the moat forced to shift "away from raw model capability - where the open-weight gap is now measured in months - toward distribution, memory and context." That is the efficiency-paradigm and good-enough theses stated by a strategist rather than by us, and it landed on the same day Anthropic committed a further US$35bn to rented capacity. Hou's "enough supply" reading is the bolder claim: it asserts capability sufficiency for most work, not merely price competition, and it is the one that would most change the investment case if true. The index is a public series, so this is testable weekly rather than annually. Watch whether it stabilises near a dollar or keeps sliding; whether frontier labs answer with price cuts - Anthropic's cache-read reduction the same day is one - or with bundling that removes per-token comparability altogether. Watch, above all, for any lab reporting tokens served or revenue per token, which is what would let the volume argument be checked instead of asserted. If pricing disclosure becomes less granular from here, that is itself an answer.