The Investor

What actually decoupled

Software was underwritten as a fixed-cost business: roughly 80% gross margins, flat per-seat pricing, engagement that cost almost nothing extra to serve. Generative features made cost of goods a function of usage, so the activity that justifies the multiple now eats the margin sitting under it. The mechanism is a year old. Pricing it into forward guidance instead of absorbing it quietly is not.

Salesforce is pushing CRM toward consumption- and outcome-based pricing as part of its agentic strategy, and the comp set follows the category-definer, which turns a margin question into a multiple question. Revenue goes usage-variable at the same moment cost does, and in most companies the cost side is unmeasured: roughly two-thirds of surveyed enterprises run AI workloads in production, many with no visibility into infrastructure cost or utilization at all. Inference does not end at go-live. It scales with the engagement the growth model celebrates.

Where the evidence pulls in opposite directions

Read as structural erosion, this argues for a portfolio-wide margin haircut. The counter-evidence sits inside the same disclosure. Canva rebuilt its architecture and reports cost per task down nearly 90% since its AI product launched in April, which reframes first-generation features as unoptimized engineering rather than economic law, and implies a competent engineering organization reclaims an order of magnitude in about four months.

The escape hatch most gross-margin plans assume, migration to cheap open weights, is thinner than the price sheets. DeepSeek prices at $0.435 input and $0.87 output per million tokens against Claude Opus 5 at $5 and $25, a decision that looks like it makes itself. Then DeepSeek pre-announced a significant increase to that API pricing, and engineer testers report reasoning-continuity failures and worse coding output than DeepSeek's own smaller model. The Information separately carries an industry analysis claiming premium OpenAI and Anthropic models finish tasks at lower effective cost (total tokens, retries, and hours to finish the job) than Chinese rivals. The sponsoring organization is undisclosed and the campaign was segmented to an Anthropic audience, so treat it as a procurement talking point rather than validation. The metric shift survives whoever paid for it. With AWS instructing its own engineers to cut CPU waste, the honest base case for hosting cost is flat to up through 2027.

The demand-side version nobody is pairing with it

Upwork cut its full-year revenue outlook with active clients down 4% to 763,000, naming AI automation, a weak labor market, and deteriorating Google search performance in one breath. Two of the three are structural for anyone selling human hours. The third generalizes across the whole book: every customer-acquisition assumption that runs through organic search needs re-testing with search decay as the base case rather than the tail.

A software company cut growth guidance because usage got more expensive, not because customers left. Underwrite gross margin per task, not ARR growth.

The line between recoverable and permanent compression is inference engineering capability, testable in four questions: cost per task, inference gross margin by feature, share of cost of goods on frontier APIs versus self-hosted or distilled models, and per-seat versus per-use pricing. The Canva answer has a bill attached, since engineering time spent reclaiming an order of magnitude is time not spent on the roadmap the multiple was underwriting. A company that cannot answer inside 48 hours does not know its gross margin, which makes its next mark wrong in a direction you cannot size.

The queue was never a demand curve

Start with the datapoint that quietly voids a great deal of spreadsheet labour: Bloomberg reports that more than two-thirds of the electricity sought for US AI data centers is unlikely to ever materialize, inflated by phantom projects and long-shot pitches parked in interconnection queues. Almost every AI infrastructure TAM of the last eighteen months used queue volume as its demand proxy, because the numbers were public and already in a spreadsheet. If deliverable capacity is a third of requested capacity, those TAMs run roughly 3x too large. That is a mark-accuracy problem on positions already held, not a selection problem on deals not yet done.

The inverse is the more interesting version. If requested megawatts are mostly fiction, whoever holds secured, energized megawatts owns something scarcer than consensus assumes, which puts three questions in the template: what share of the pipeline carries a signed power purchase agreement with a named offtaker, what this specific utility's historical queue-attrition rate is, and what the energization date is and who eats the cost of delay. Duration says the same thing from the equipment side. IBM's $240M multi-year Nvidia HGX B300 commitment for Together AI does not come available until Q1 2027, which makes on-demand GPU assumptions inside 2026 operating plans closer to fiction than forecast.

What the tape actually paid for

The table reads as one sentence: the market paid for the unverifiable capacity narrative and charged the two companies that disclosed granular revenue mix. Cerebras' $25.4B backlog looks enormous until you notice it moved $400M in a quarter, a significant share of it belonging to one customer. Second print since May's IPO.

Cerebras disclosed its customer concentration and lost 16%. Nebius asserted it could sell 2027 capacity and gained 17%. Equilibria that pay for opacity do not last.

Two source conflicts worth carrying

Bloomberg's market snapshot has Cerebras up 12% while its own accompanying headline reports shares tumbling on guidance, so treat intraday direction as unverified until the tape confirms it. Bloomberg also cites Nebius cloud sales up 514% in Q2 while The Information puts capex at 9.7x quarterly revenue. Both are true and not in tension: the growth is real, and the question is who funds it. Morning Brew's read-through, CoreWeave closing at $107.73, up 19.28% in a session alongside Supermicro and Nebius beats, is the public comp set neocloud marks get struck against, and comp windows in AI compute stay open for weeks.

Underneath sits the circularity nobody prices as a discount. Nvidia and six financial firms are standing up financing pools reported at roughly $500B, while OpenAI and Anthropic are projected to spend $200B+ on 2027 compute funded by a $250-300B raise. Either the demand is genuine and the financing merely accelerates it, or the financing is manufacturing demand that would not otherwise clear. This is probably the second, held loosely. When the supplier finances the customer, demand and capital stop being independent signals, and any memo citing Nvidia participation as a quality marker is mispriced by construction. Take the haircut.

Why calendars are pricing companies

Cohort pricing is not laziness, or rather it is the laziness that shows up when comparable price discovery disappears. Megadeals took 87.5% of US venture dollars in the quarter, which leaves very few arms-length marks on ordinary companies. A buyer with an LP clock and no clean comparables defaults to the one variable that is always public and never ambiguous: the year of the last round. Every 2021 mark is presumed fiction, every 2026 mark presumed current, and the years in between get skipped.

The counter-thesis deserves airtime because it is mostly right: 2021 valuations were set in a different rate regime and never marked honestly afterward. True of most of the cohort, not all of it, and separating the two is the entire job. What keeps the dislocation finite is that vintage-based pricing does not survive contact with a disciplined buyer. Price one 2021-vintage company off measured unit economics and the cohort discount on that name stops existing.

A forced seller cleared a record

Set the secondaries tape against a record-setting private transaction. The Lakers changed hands at $12.5B, roughly 25% above the ~$10B Mark Walter paid the Buss family twelve months earlier, to a group led by Bob Iger and Thrive Capital's Josh Kushner. Per Morning Brew's account of Bloomberg's reporting, Walter is under federal investigation for alleged tax fraud involving billions in loans to his companies, and the sale expedites his effort to raise cash. A liquidity-constrained seller under legal pressure took a premium rather than a haircut.

This is probably too tidy, but put the two facts side by side and the read is not about liquidity. Illegible assets with contested marks take a cohort discount; legible, scarce assets with observable cash flows absorb forced supply at a record. The variable being priced is mark legibility, the same variable the public tape used when it docked a chipmaker for disclosing customer concentration. Thrive is building the second category on purpose: Thrive Eternal bought into the San Francisco Giants and pursued a World Cup stake before FIFA scrapped its private-investment plan.

Trophy assets clear at a premium from a distressed seller while 2021 venture paper clears at 59% off. The discount is not paying for illiquidity; it is paying for marks nobody can verify.

Two caveats before that comp reaches a valuation committee. The Lakers transaction requires league approval, and the buyers carry real optics risk: Iger ran Disney, an NBA broadcast partner, until March 2026, and Kushner's brother is an unofficial White House advisor. FIFA's reversal established that stakeholder sentiment can veto private capital in core sports properties. Separate durable value (market, media rights, brand, real estate) from roster-contingent value built around one superstar, and haircut the latter hard.

The operable version for the book is narrow and cheap. Rank the 2020-2022 holdings by whether they can produce audited per-unit economics and customer concentration on two weeks' notice, which costs a fortnight of analyst time otherwise spent screening new deals. Those are the names the cohort discount misprices, and the only ones you can defend to an LPAC when a calendar year is doing the pricing.