ClarityBeta

Leadership & Executive

The Board Room

9 min read · 1,815 words · 3 moves to watch
The Signal

AT&T cut its AI bill 56% in the same month chip rental renewals repriced 20% higher.

The quality cost came published alongside the savings: 2% degradation, measured. That gives you a number to hold a vendor to at the next renewal, which is more than most model contracts currently contain. Capacity, meanwhile, is allocated by cash position: Oracle has booked $11.4B in customer prepayments and OpenAI has stopped selling its Pro tier outright, so netting savings against supply risk into a single AI line misprices both sides of the plan.

In Play

  1. Open Weights Took the Lead as Enterprise AI Spend Contracts

    AT&T cut its AI bill 56% across a 100,000-employee estate and measured a 2% quality decline, per The Pragmatic Engineer's reporting on this month's enterprise disclosures. On log10.io's clinical-regulatory benchmark, open-weight GLM-5.3 now ranks first on the hardest agentic task at 88.2 against GPT-6 Astra's 86.2. Cost per completed run across the top ten spans $0.30 to $22.29, and Opus 5 sits at the top of that range while scoring below the $0.30 model.

    Ask Clarity
    Try

  2. Compute Access Repriced Upward and Prepay-Only

    Oracle funded $28.5 billion of capital expenditure while burning only $5 billion of its own cash, closing the gap with $11.4 billion of customer prepayments. Co-CEO Clay Magouyrk disclosed that Nvidia chip rental renewals came in at a 20% premium to prior contracts. OpenAI stopped selling its $200-a-month Pro tier because it could not serve demand. Capacity is being allocated by cash position, which pulls your treasury function into AI strategy.

    Ask Clarity
    Try

  3. Agent-Mediated Access Breaks Seat Pricing

    At Goldman Sachs' technology conference in San Francisco, Salesforce told investors 'no browser required: the API is the UI.' Figma's Dylan Field took the opposite position, and Figma is down 40% year to date while guiding to slower September-quarter growth. HubSpot said more than 350,000 workers already reach it through external AI tools. Any revenue line priced on the assumption that a human opens your product is now exposed.

    Ask Clarity
    Try

  4. Product Quality Stopped Working as a Moat

    Meta shipped Muse, its own AI assistant, roughly a week after Instinct reportedly turned down a 10-figure offer, per Not Boring. Instinct had raised $350 million at a $2.5 billion valuation and is the category's acknowledged best product. Copying it costs Meta nothing, so being better bought no defense. The card market shows what does: Ramp, which earns when customers spend less, is worth $44 billion; Brex and Divvy played the incumbents' points game and sold for $5.15 billion and $2.5 billion.

    Ask Clarity
    Try

  5. 2021 Valuations Are Being Marked as Information Gaps

    Forge's August report showed the median secondary trade pricing at par to the last primary round for the first time since early 2022. The composition explains it: companies whose last priced round was in 2021 still trade 59.1% below it, against 0% for 2026 vintages. Airtable agreed to sell to Bending Spoons at roughly $2.25 billion, about 81% below its 2021 mark, with around $480 million of ARR growing over 20%. Operating performance no longer defends a stale price.

    Ask Clarity
    Try

Deep Dives

Inference Deflated, Capacity Inflated — and Only One Shows Up in Your Renewal

Two AI cost curves are now moving in opposite directions, and any leader who nets them into a single line item will underprice the compute plan and overpay the model bill.

The two curves land on different lines of your P&L

Substitutable work is getting cheaper fast. Uber has held total AI cost flat since March while token and session volume kept climbing, with cost per request down 34% and cost per session down 52%, per The Pragmatic Engineer's account of the company's disclosures. On Uber's own numbers, its most expensive open-weight model runs $0.30 per code review against $0.50 for the cheapest frontier model and $2.50 for the most expensive. SpaceX reportedly cut Claude Code tokens by 90%. None of that required new science; it required a gateway between the application and the model plus a weekly benchmark.

Scarce capacity is moving the other way. Oracle's quarter, as reported by The Information, is a financing event rather than a demand event: management confirmed that most of the additional $30 billion of AI compute deals booked in the period came through customer prepayment or bring-your-own-chip structures. The market has quietly reorganized so that customers capitalize their vendor's build-out.

Capacity procurement is now a treasury decision

StructureCash timingUnit cost trajectoryCounterparty exposure
Traditional rentalPay as consumedWorst — repricing at +20% on renewalLow; walk away at term
Prepaid capacityLarge upfront outlayBest — you lock today's priceHigh; your cash sits on their balance sheet
BYO-chip colocationCapex now, low opex laterInsulated from rental premiumsMedium; hardware is yours, siting is not

Note the trap in row two. Prepayment is the only structure that reliably caps unit cost, and it maximizes exposure to a vendor S&P downgraded in July over OpenAI concentration and capex, whose stock is down 53% over twelve months and which is actively managing data center delays. Magouyrk's line that "all our eggs are not in a single basket" is site-diversification messaging, not a delivery guarantee.

What the 75x spread does not include

The log10.io figures are verified list prices per run and exclude self-hosting, GPU capacity, ops headcount, model validation, and regulated qualification work. The loaded multiple for a self-hosted open-weight deployment lands well below 75x and is still large enough to reorder a budget — but nobody publishing a number knows where. Two further constraints travel with the benchmark: the 0.6-point gap between first and second place is inside judge-panel noise, and the two labs driving the open-weight surge are Chinese-origin, with export-control and customer-contract review nowhere in the analysis. The enterprise savings figures, separately, are self-reported by the companies claiming them.

Where the evidence agrees, and where it diverges

Every credible account reviewed converges on the same architectural answer: a provider-agnostic routing layer, bought rather than built for version one, with an explicit quality-regression budget — AT&T's measured 2% is a defensible ceiling. The divergence is the interesting part. Ramp's telemetry shows AI spend falling 10% in August at the largest buyers, while Oracle books record AI bookings and OpenAI refuses revenue for lack of capacity. Both are true, and together they describe the actual market: buyers are driving down what they pay per unit of routine work while paying more for the scarce top tier.

The frontier premium is now a negotiable line item, and the only part of the compute bill that is still rising is the part you cannot substitute.

What to do

  1. Rebuild the three-year compute forecast at flat-to-plus-20% unit pricing and publish a per-SKU AI contribution margin table with breakeven pricing before the next board cycle.

  2. Reopen frontier model contracts ahead of renewal using the published AT&T and Pinterest figures as pricing anchors, and refuse any commitment past twelve months without a benchmarking or price-adjustment clause.

  3. Commission one fully loaded self-hosted open-weight TCO — capacity, validation and qualification included — alongside an export-control and customer-contract review of Chinese-origin weights, this quarter.

The Only Moat Left Is a Price Your Competitor Cannot Match

Product quality and the interface both became copyable this month; what survives is a monetization model an incumbent would have to break its own P&L to imitate.

Figma is the most agent-exposed and the best priced for it

The three CEO positions read as one spectrum, and the spectrum contains an arbitrage. Salesforce co-founder Parker Harris asked publicly, "Why should you ever log into Salesforce again?" That is a comfortable question to ask, because agents still traverse CRM records to do anything useful, and executive Bill Patterson claims reaching Salesforce through Claude is increasing the work flowing through the system. Salesforce still has no answer for how it charges when an external agent touches that data. Figma sits at the other end of the spectrum. A general-purpose chatbot can produce design output without touching Figma at all, and the hybrid subscription-plus-consumption model is the only reason Dylan Field could call chatbot dominance "not entirely bearish."

VendorWhere value sitsAgent exposureCan it bill for machine access?
SalesforceSystem of recordLowUnsolved, openly
HubSpotRecord plus deep operational workflowMediumClaims agent access lifts engagement and retention
FigmaThe interface itselfHighBest in class — consumption already captures it

The most exposed company has the best pricing architecture. The least exposed has the weakest one and says so in public. Salesforce's next pricing decision has to name a billable unit of agent access before the usage growth Patterson describes shows up as revenue.

The persona split is the real roadmap input

HubSpot's Yamini Rangan gave the most operationally useful frame of the week. A campaign manager running a multi-region campaign against 10 million contacts "is going to work in HubSpot all day long," while a CMO preparing for a meeting pulls the artifact through an AI interface. Depth stays in the app. Shallow reads migrate. Applied to a product roadmap, that defunds the middle: the casual-user dashboards built so occasional visitors feel served. Occasional visitors will not return to those dashboards; their agents will read the data instead.

What the corporate-card endgame proves about defense

Three challengers in one market over the same period finished with roughly an 8x valuation spread, and the outcome was decided years before it printed. Ramp made money when customers spent less, so matching it meant incumbents reporting lower revenue to shareholders. The shareholders would not authorize that trade. Brex and Divvy played the incumbents' own points game slightly better and were acquired. Instinct is the live version of the same test: the acknowledged best product in the hottest category, with no revenue line anywhere that breaks when Meta copies it. For Meta, copying is better than free, because its user graph solves the cold-start problem that is Instinct's hardest engineering challenge.

Seat pricing is the 2026 handcuff

Every immobilized incumbent in the record was disabled by its best revenue line. The 2026 version of that handcuff is seat-based and consumption-linked pricing. An agentic entrant billing on outcomes is counter-positioned against an incumbent exactly as Ramp was against the points issuers. The product is visible and buildable in a quarter, and it cannot be matched on price without shrinking revenue per customer. That is not a product gap. Shipping faster will not close the gap, because the gap is in the price sheet.

One caveat on the vendor claims. Every assertion that agent access grows usage comes from executives with an interest in calming investors. Treat "better engagement and better retention" as a hypothesis to verify against internal telemetry.

Incumbents could see exactly what Ramp was doing and still could not match the price without reporting lower revenue.

What to do

  1. Run a two-week agent-mediated revenue exposure audit, owned jointly by the CFO and CPO: what share of ARR assumes a human logs in, and which of those events are technically meterable today.

  2. Ship metered, entitlement-aware agent access to your core data this quarter — per-call billing, rate limits, and governance logging — before customers demand it.

  3. Commission a CFO-owned memo naming the high-margin line that prevents you from adopting an outcome-based price, with the response trigger pre-authorized this quarter.

A Valuation You Haven't Refreshed Since 2021 Is Priced as an Information Problem

The secondary market's cheerful headline hides a vintage split that now governs your recruiting story, your financing window, and what you can buy while everyone else is discounted.

The distribution, not the median

Forge's median sat at par in June and 7% below in July, which is the number that gets quoted and the least useful one available. The 25th percentile sat at a 34% discount and the 10th percentile widened from 50% to 57% in that same month. One name drove the tape: SambaNova's Forge Price rose 142.9% in July and contributed 6.5 of the 9.3 percentage points of gain in Forge's equal-weighted index. The model ingests primary-round pricing. That is a repriced round, not an inflecting business.

SambaNova reset the denominator on purpose

April 2021: Series D at $5.1B. December 2025: Bloomberg reports Intel near a deal at about $1.6B; talks stall. February 2026: a $350M Series E at a reported $2.2B. July 2026: a $1B Series F first close at $11B post-money, led by General Atlantic with T. Rowe Price, Capital Group, BlackRock and QIA in the syndicate. The February markdown gave buyers a recent reference to price against, which is what made the July raise possible.

Reference roundImplied valuationHow the same $2B secondary trade reads
Series D, April 2021$5.1B61% discount — stale mark
Series E, February 2026~$2.2B9% discount — healthy name
Series F, July 2026$11B82% discount — reads distressed

The secondary trade price is the constant across that sequence; the reference round is the variable. Taking the down round was the cost of holding a current mark.

Capital access itself has become the signal

PitchBook's Emily Zheng put the mechanism plainly: "Companies that cannot raise on strong terms right now generally are not raising at all." With 86% of H1's $412.7B going to AI and 87.5% into $100M-plus megadeals, the old posture of raising when ready now reads as inability, and silence is priced at 50 to 60% off. The liquidity underneath is thinning at the same time. Buy-side indications fell to 48% in July from 57% in June, the first non-majority month since late 2023, while the 90th-percentile premium to last round collapsed from 79% to 27% and the 75th from 23% to 7%. More than 220 former unicorns now sit below $1B, 75 of them SaaS.

The same window makes capability cheap to buy

Bending Spoons is acquiring Miro at $1.355B against a $17.5B peak, a 92% discount, and a public comp for the entire 2021 collaboration cohort. A skeptic would say one distressed sale does not set a comp, and the skeptic is right up until an acquirer cites it in a term sheet. The tradeoff is explicit: capability that would otherwise cost eighteen months and a hiring cycle is available at low-single-digit revenue multiples, and the competing bidders for those assets are cost-focused roll-up operators, not strategic buyers. Volition closed a record $950M fund writing $25–50M checks into $5–50M-revenue companies, and its GP reports seeing more companies reach $10M ARR inside year one with fewer than 10 employees than ever before, against a historical bar where $5M ARR implied years of operation. The comparison set for the next round is a sub-ten-person company at $10M ARR.

Two caveats to hold: Forge discloses its price "may rely on a very limited number of trade and/or IOI inputs" with undisclosed weightings, and the sub-10-person cohort claim is one general partner's observation, not market data.

Carrying a stale mark costs roughly 59% when someone finally prices it.

What to do

  1. Ban the headline secondary median from board and compensation materials and replace it with your own vintage-cohort discount and full percentile distribution before the next board meeting.

  2. Decide explicitly this quarter whether to reset your reference price — insider-led round, structured extension, or company-run tender — if your last priced round predates 2023.

  3. Rank five acquisition targets at sub-2x revenue and formally reverse one queued build decision to buy this quarter.

The bottom line

One pattern runs through today: the things you could once buy your way to — better models, better product, more engineers — are converging toward commodity prices, while the things you cannot substitute are being rationed by cash and by contract. That inverts the assumption most operating plans still rest on, that scale and spend buy protection; from here, protection comes from what you own outright and what your competitor cannot charge for without hurting itself. Split the AI plan into two books this quarter — one for capability now purchasable at commodity rates, one for the capacity, data and pricing position you must secure ahead of need — and give a single named executive the boundary between them.