AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #97

Buyers Set The Terms

Tuesday, August 25, 2026 · 4 Developments · 2 Perspectives · 1 Case Study · 2 Resources

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

For two years the labs set the terms and buyers adjusted. Today's items all point the other way, and the shift is showing up in prices, in build decisions, and in what enterprises are willing to certify before an agent touches production.

The through line: leverage moved to the buyer, and the companies capturing it are the ones that measured first.

Developments

CompetitionDeveloping

August 24: Anthropic's flagship took 11% of its own corporate spend, and OpenAI reset frontier prices

The Financial Times reported on August 24 that Fable 5 accounts for roughly 11% of enterprise spending on Anthropic models two months after launch, using Ramp payment data across about 70,000 companies. Anthropic's cheaper Opus 5 has passed it. Days earlier OpenAI cut GPT-5.6 Sol by more than 20% for three months, to $4 per million input tokens and $20 per million output, undercutting both Anthropic tiers. A16z partner Martin Casado offers a non-price explanation: Fable retains customer data and ships no zero-retention option, which disqualifies it at companies with strict data policies.

So what

Two separate buying constraints are now visible in the same data, and only one of them is price. Capability at the top of the range is no longer the deciding factor; cost and data handling are. If you sell software with AI inside it, your procurement conversations are about to include retention terms, not just per-seat pricing. Publish your data-handling posture on your pricing page this quarter, and check whether your own model contract lets you make the promise your buyers will ask for.

Hype check

Several outlets rendered the 11% as a share of all corporate AI spending. It is the Fable share of identified Anthropic spend within one payments platform's customer base, which is a narrower and less alarming claim.

Confidence: Established fact — the pricing change is published by OpenAI and the spending figures come from Ramp data reported by the FT and corroborated across outlets.

Go deeper Financial Times · The Deep View

Technology

August 24: Nvidia's first standalone CPU targets the part of agent work nobody budgets for

Nvidia announced that SpaceXAI will deploy its new Vera CPU to run Grok's agentic systems, and will fly a space-optimized version aboard its Starmind satellites. Vera is aimed at the orchestration layer rather than inference: running code, calling tools, and querying databases between model turns. Nvidia claims up to 1.8x faster agentic task completion than x86. The company has never sold standalone CPUs before and its CFO has put the addressable market at roughly $200 billion.

So what

The interesting signal is not the satellite. It is that the largest supplier in AI infrastructure has concluded the binding constraint on agents is coordination overhead, not model quality. If your agent costs are running above plan, the likely cause is idle time and retries in the orchestration loop rather than token price. Instrument wall-clock time per completed task alongside token spend before your next capacity decision, because the two numbers are drifting apart.

Hype check

The performance figures are Nvidia's own and unaudited, and analysts put orbital compute at more than four times the cost of ground compute today. Treat the satellite timeline as a launch-customer arrangement, not a product you can buy.

Confidence: Established fact — the partnership and chip specifications are announced by Nvidia, though the benchmark claims are vendor-supplied.

Go deeper Nvidia Newsroom · The Rundown AI

Technology

August 24: Chinese state-linked crews doubled attack volume after adopting DeepSeek

Taiwanese research firm TeamT5 told Bloomberg that tracked Chinese state-tied groups are running more than twice as many attacks as they did before folding open-weight models into their workflow. The groups favor DeepSeek specifically for its low cost and weak guardrails, and researchers have not yet observed pricier models such as Kimi K3 in an attack. Documented uses include writing intrusion code, raiding a Taiwanese firm's email, and mapping a target across a thousand addresses. The UK AI Security Institute warned in May that AI cyber capability is doubling every few months.

So what

Most security planning still assumes the dangerous capability sits behind a frontier lab's safety review. The evidence points the other way: the cheap downloadable model with no guardrails is the operational one. Attack volume, not attack sophistication, is what changes your exposure math. Ask your security lead to model a doubling of inbound phishing and reconnaissance against current detection capacity, and treat the answer as a 2027 budget input rather than a tabletop exercise.

Confidence: Informed perspective — the finding comes from one research firm's tracking of a subset of groups, reported by Bloomberg but not independently replicated.

Go deeper Bloomberg · AI Security Institute

Business

Nvidia is in talks to back Perplexity at $30 billion, roughly 40 times run-rate revenue

The Information reported that Nvidia is discussing an equity investment in Perplexity at a valuation as high as $30 billion, with a technology licensing arrangement also under consideration. Perplexity's revenue has more than tripled this year to above $750 million on an annualized basis. Nvidia is already an investor alongside Jeff Bezos and SoftBank. Terms are unconfirmed and neither company has commented.

So what

Forty times run-rate revenue is the number your board will hear quoted, and it is set by a strategic buyer with a supply-chain motive rather than a financial one. That distinction matters when you benchmark your own round. Growth rate, not multiple, is doing the work here: revenue tripling in eight months is what supports the price, and a comparable multiple without comparable growth will not survive diligence. Anchor your next raise on the growth curve you can evidence, not on the headline someone else printed.

Confidence: Watch this — the talks are reported by a single outlet with no confirmation from either party and no closed terms.

Go deeper The Information · The Rundown AI

Perspectives

AI Strategy & Transformation

Public benchmarks cannot tell you whether a model does your company's work

Mercor CEO Brendan Foody says he regularly meets executives spending up to $100 million a year on models with no offline evals, meaning no fixed set of real tasks used to compare candidates before they touch live work. Box CEO Aaron Levie summarized the problem as enterprises no longer being able to buy on vibes. Every describes its own answer: an internal eval built from about 30,000 past edits by its editor in chief, which tracks which suggestions editors accept and what they still rewrite afterward. That residual rewriting is the number public leaderboards never show.

So what

With open models taking real volume and prices moving every few weeks, the ability to swap models cheaply is becoming a cost lever worth more than any single negotiation. You cannot swap what you cannot measure. Pick one recurring job this week, write down five ways your current model gets it wrong with one gold-standard example each, and run every candidate against those cases. That artifact will outlast three model generations and tells you more than another afternoon on leaderboards.

Confidence: Informed perspective — practitioner argument supported by named executives and one company's disclosed internal method, not by a survey.

Go deeper Every · Brendan Foody

Operations & Finance

Agent rollouts are stalling in risk review, and the teams getting through share three habits

Sam Altman has attributed slow agent adoption to economic inertia. The more specific explanation is that 66% of companies name security and risk as their top barrier to scaling agents, per McKinsey's 2026 State of AI Trust. Reporting on the organizations getting past that gate finds a consistent pattern: they replace end-to-end workflows with single-responsibility agents, require a written log of every decision and action, and insert human review before any high-stakes execution. The constraint is what legal and compliance will certify, not what the model can do.

So what

Narrow scope is usually treated as a limitation. Here it is the thing that gets the project approved, which makes it a speed advantage rather than a compromise. If an agent initiative has been sitting in review for a quarter, the fix is almost certainly decomposition rather than a better model. Take your stalled workflow, split it into the three narrowest agents that each own one reversible step, and put the audit log in front of your risk lead before you ask for the next approval.

Confidence: Informed perspective — the McKinsey barrier figure is survey data, while the three shared traits are a pattern drawn from reported cases rather than a controlled study.

Go deeper VentureBeat · McKinsey

Case Studies

AI Strategy & TransformationDeveloping

August 24: Thomson Reuters spent $40 million to stop renting its own intelligence

Thomson Reuters launched Thomson, a proprietary model built by starting from an open-weight base and training on decades of Westlaw, Practical Law, Checkpoint, and Reuters content. Two years cost about $40 million in staff and compute; the most recent training run cost $450,000. Only about 10% of the company's content library has been used so far. CTO Joel Hron compared renting a model to renting a house, with no equity that compounds. It ships first inside CoCounsel Legal, which stays multi-model.

So what

Forty million dollars over two years is roughly a year of API invoices at enterprise scale, which puts owning a domain model inside reach of companies far smaller than Thomson Reuters. The gating asset is proprietary content that is already structured and rights-clear, not a research team. Have your team price two years of projected inference against a scoped fine-tune on your own corpus. If the numbers are close, the deciding question is whether your data is genuinely differentiated, and that is a conversation to have before your next multi-year vendor commitment.

Hype check

Several outlets name Alibaba's Qwen as the base model, but Thomson Reuters' own launch materials do not identify it. The benchmark comparisons against frontier models are also company-run and internal so far.

Confidence: Established fact — the launch, cost, and architecture are disclosed by the company, though the performance claims are self-reported.

Go deeper Thomson Reuters · The Deep View

Resources

Product & Technology

A model that passes 65% of single attempts held up on only 25% across twenty runs

Two new papers give leaders better questions to ask about agent reliability. Thinkingbox found that the strongest coding model tested passed 65% of single attempts, but its success rate fell to 25% when researchers required it to succeed across twenty consecutive attempts on the same task. CentaurBench found that the model best at completing a task alone was frequently not the best at improving a weaker model's first draft, with a different model winning on five of seven tasks.

So what

Every agent demo you have been shown is a single attempt, and single-attempt accuracy is the wrong metric for anything that runs unattended in production. Consistency across repeated runs is the number that determines your support load. Add a pass-at-twenty check to your evaluation criteria for any agent you intend to run without a human in the loop, and pair the two findings: the model you use to generate may not be the one you should use to review.

Confidence: Established fact — both findings come from published preprints with stated methods, though neither has been peer reviewed.

Go deeper Thinkingbox (arXiv) · CentaurBench (arXiv)

People & Organization

A six-question framework for people whose roles are being redrawn

Cliff Maxwell and Jobs to Be Done co-creator Bobby Moesta published a guide built from more than 1,000 interviews with professionals changing jobs. The core distinction is between progression, the ladder model that AI is dissolving as job boundaries blur, and progress, which is contextual and can look lateral from outside. The practical tools are a pushes-and-pulls exercise, four common motivations for change, and an energy profile built by auditing a month of your calendar for what energized rather than drained you. The World Economic Forum projects 39% of job skills will change by 2030.

So what

Your retention risk in the next eighteen months is concentrated in strong performers whose jobs are quietly changing shape underneath them, and most will not raise it until they have already interviewed elsewhere. The energy-profile exercise works as a manager conversation, not only a job-search tool. Ask your top ten people which parts of their current work energize them and which drain them, then redesign around the answers before AI redraws the roles for you.

Confidence: Informed perspective — a practitioner framework grounded in the authors' own interview base, with the skills-change figure sourced to the World Economic Forum.

Go deeper Lenny's Newsletter · World Economic Forum

Sources in this edition: The Neuron, The Rundown AI, The Deep View, Superhuman AI, Every, Lenny's Newsletter, TLDR AI, Daily AI

← Latest Editions and Search