Most of today's items describe the same move made by different companies: deciding to own a layer they used to rent. OpenAI built its own chip, Perplexity and Apple pushed inference onto hardware you buy once, Google absorbed the vertical workflow, and a third of McKinsey's respondents skipped a software purchase and built the feature themselves.
- Local inference ships. Perplexity, Nvidia, and Apple all launched on August 25 for the buyer who wants to stop paying per token.
- OpenAI's own silicon. Jalapeño beat Nvidia's Blackwell on performance per watt in an open benchmark, under two years from concept.
- Google goes vertical. Gemini Enterprise now ships tuned editions for legal and financial services, with four large law firms named at launch.
- Agents get your credentials. ChatGPT Work can now sign in to the systems your employees use and finish the task itself.
- A $30 trillion pitch. Anthropic will reportedly tell IPO investors its addressable market dwarfs the combined revenue of corporate America.
- Price, not consumption. Enterprise software spend rose 8% while app counts fell, and 78% of IT leaders got surprise AI charges.
- The byline question. A billionaire investor said he used AI to write his newspaper column, and his publisher defended him.
- Judgment as a skill. Every turned 30,000 of its editor's past edits into an agent anyone on staff can run before filing.
- Six percent, not thirty-seven. McKinsey found material EBIT impact at 6% of firms, unchanged for a second year.
- Credentials that expire. Vercel made task-scoped, short-lived agent credentials generally available with more than 100 connectors.
If your product sits on a layer a competent buyer could now assemble themselves, that is the sentence to bring to your next board meeting.
Developments
Technology
August 25: Perplexity, Nvidia, and Apple all shipped for the buyer who wants to stop paying per token
Perplexity and Nvidia launched Portable Computer, a version of Perplexity's agent that runs entirely on a user's own machine with no billing credits, starting on Nvidia's $4,699 DGX Spark. It runs PPLX 27B, a post-trained Alibaba Qwen 3.8 27B, and asks permission before sending any step to a cloud model. Linux ships now, Windows in September, and users need an RTX GPU with at least 24GB of VRAM. Hours earlier Apple introduced the M6 Mac mini at $899 with what it says is 4x faster AI performance, plus an M5 Ultra Mac Studio at $5,499, both shipping September 22.
So what
A one-time hardware purchase between $900 and $5,000 now covers a meaningful share of agent work that used to be metered, and heavy users were already spending twice that per month on tokens. That changes two things at once: your own cost line, and your customers' willingness to accept usage-based pricing from you. Run the arithmetic on your three highest-volume internal workflows against a fixed-cost local box this quarter, and separately audit which of your product's AI features a customer could now replicate on their own hardware.
Confidence: Established fact — both launches are announced by the companies with published pricing, availability, and hardware requirements.
Go deeper Perplexity · Apple Newsroom · The Deep View
Technology
August 25: OpenAI's first custom chip beat Nvidia's Blackwell on efficiency, and it used its own models to design it
At the Hot Chips conference OpenAI published first results for Jalapeño, an inference-only accelerator built with Broadcom. Tested on SemiAnalysis's open-source InferenceX benchmark using GPT-OSS, DeepSeek R1, and Kimi K2.5, the 700-watt part delivered 1.5 to 4 times the performance per watt of the leading commercial systems, including Nvidia's 1,200-watt Blackwell. OpenAI will not sell it. The company says its own Astra and Codex models helped optimize the circuits and kernels, taking the design from concept in late 2024 to working silicon, with volume ramping through 2027.
So what
The durable news is not the benchmark, it is the timeline: a software company reached production silicon in under two years by using its own models on the hardest part of the work. Every large lab now owns silicon, which means inference prices over the next two years will be set by internal cost structures you cannot see. Stop treating published API prices as a stable planning input, and negotiate committed-use terms with a floor rather than assuming the next cut arrives on schedule.
Hype check
The benchmark is open source and the model mix was not OpenAI's alone, but OpenAI ran the tests and chose the comparison points. No independent lab has reproduced them, and the chip has not been measured against dedicated inference rivals such as Groq or Cerebras.
Confidence: Established fact — the chip and its results are published by OpenAI and characterized independently by SemiAnalysis, though the tests are vendor-run.
Go deeper OpenAI · SemiAnalysis · The Register
Business
August 25: Google put Gemini Enterprise inside legal and financial services workflows
Google Cloud released its first industry-specific editions of Gemini Enterprise, in preview for legal and for financial services, with healthcare and life sciences named as next. The legal edition ships pre-built agents and skills for brief drafting, citation verification, contract lifecycle management, regulatory scanning, and data subject access requests, plus connectors into the document management, e-discovery, and research systems firms already run. Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly are named launch firms. Weil has already built a judicial-insights product on top of it.
So what
A hyperscaler is now selling the workflow layer, not the model layer, and it arrived with connectors into the systems of record that vertical software companies have long treated as their moat. If you sell into legal, financial services, healthcare, or life sciences, assume a Google-packaged competitor appears in your deals within four quarters. The defensible ground is proprietary outcome data and workflow depth that a horizontal agent cannot infer, so name yours explicitly in your next board deck and fund it.
Confidence: Established fact — announced by Google Cloud with named launch customers, though the products are in preview and adoption is unproven.
Go deeper Google Cloud · Artificial Lawyer
Technology
August 26: ChatGPT Work can now sign in to the systems your employees use
OpenAI extended ChatGPT Work so its agent can authenticate to websites and carry on with a task instead of handing control back to the user. OpenAI says the model never sees the credentials, that they are not stored or used for training, and that the agent asks for confirmation before consequential actions. Reporting differs on the mechanism: some accounts describe password-manager style autofill, others describe the agent surfacing the login screen for the user to complete. It is rolling out on web and mobile for Plus, Pro, and Business accounts.
So what
Agents that hold sessions inside your SaaS applications produce activity your access logs will attribute to a named employee, which breaks the audit trail every SOC 2 and enterprise security questionnaire depends on. This will land on your desk as a customer question before it lands as an internal policy. Decide now whether agent sessions get their own service identities with scoped permissions, and put that answer in writing before a buyer's security team asks you for it.
Confidence: Established fact — the capability is announced by OpenAI and covered independently, though outlets describe the credential-handling mechanism differently.
Go deeper Neowin · Superhuman AI
BusinessDeveloping
August 25: Anthropic will reportedly tell IPO investors its market is worth more than $30 trillion
The Wall Street Journal reports that Anthropic's IPO materials will put its total addressable market above $30 trillion, exceeding the $28.5 trillion SpaceX floated in its May filing. The figure is derived from the estimated value of all work AI models could perform, not from software budgets. Anthropic's own 2028 revenue projection, reported earlier by Reuters, is $190 billion to $200 billion, against roughly $47 billion of current annual run-rate revenue and $11.6 billion booked last quarter.
So what
This is the labor-substitution TAM argument written into a public filing for the first time at this scale, and if it clears the SEC it becomes the template every AI vendor uses on your procurement team. Expect vendors to justify price increases by pointing at your payroll rather than your software line. Prepare the counter now: for any AI tool above a material threshold, require the vendor to show cost displacement in your own operating data before renewal, not category-level economics.
Hype check
The 1,500 largest US public companies earned roughly $2.4 trillion in combined profit last year. A $30 trillion market is a modeling exercise about the theoretical value of human work, not a revenue pool anyone has measured. Anthropic would need to grow more than 600 times to touch a single-digit share of it.
Confidence: Watch this — the figure comes from reporting on unfiled IPO materials, and Anthropic has not published it.
Go deeper Wall Street Journal · PYMNTS
Perspectives
Go-to-Market
Software prices rose four to six times faster than inflation, and only 28 cents of each new AI dollar is fresh budget
SaaStr argues the seat-based model is finishing, and the supporting data is the interesting part. Software prices rose 12% to 16.4% through 2026 against general inflation near 2.7%. The average enterprise now spends $55.7 million a year on software, up 8%, even as the number of applications in use slipped, meaning every dollar of growth came from price. Among IT leaders, 79% saw a renewal increase and 78% received unexpected AI or usage charges. Of 141 CIOs surveyed, 45% fund AI from existing software budgets and 54% are cutting vendor counts.
So what
If roughly three quarters of AI spending is funded by cutting other software, then your renewal is competing against your customer's AI roadmap, not against your named competitor. Vendor consolidation at 54% means the question in the room is which three tools survive, not whether yours is good. Get ahead of it by pricing on a value metric your buyer already reports to their own board, and by giving procurement a defensible reason to keep you when they run the consolidation exercise.
Confidence: Informed perspective — the argument is one practitioner's, though the underlying price and budget figures come from named industry surveys.
Go deeper SaaStr · TLDR Founders
Business
August 25: A billionaire investor said he used AI to write his op-ed, and his publisher defended him
Stanley Druckenmiller confirmed he used AI to write his Wall Street Journal column criticizing Treasury bond buybacks, saying the ideas were his and that he had argued them for fifteen years. WSJ opinion editor Paul Gigot backed him, framing the test as whether the argument is the author's own and whether the author has standing to make it. The Financial Times has banned its columnists from using AI. Every's Mike Taylor took the permissive side, noting that a ban would silence people with expertise who cannot write.
So what
Two credible institutions reached opposite conclusions in the same week, which tells you there is no settled norm to inherit and that your company will be judged against whichever standard your audience assumes. The exposure is concrete: investor letters, customer commitments, and regulatory filings that carry a named signature. Write a one-page disclosure standard for externally published material under an executive's name, distinguishing drafting assistance from authorship, and circulate it before someone asks you about a specific document.
Confidence: Established fact — Druckenmiller's confirmation and the WSJ statement are on the record, though the norms question remains open.
Go deeper NOTUS · Every
Resources
AI Strategy & Transformation
August 25: McKinsey's 2026 survey found real earnings impact at 6% of firms, and a third building instead of buying
McKinsey surveyed 1,719 leaders for its State of AI report. Some 37% attribute at least some EBIT impact to AI, about the same as last year, while only 6% clear the firm's high-performer bar of at least 5% of EBIT plus a self-described significant impact, also flat. Meanwhile 80% of individual users report personal productivity gains, 40% of companies above $1 billion in revenue say they are scaling agents against 27% last year, and 20% say AI operating costs have constrained usage. Nearly a third decided against buying at least one software product in favor of building it with coding agents.
So what
The build-instead-of-buy number is the one that should move your planning, because it is a demand-side statistic dressed up as an adoption statistic: roughly a third of your addressable market considered replacing a purchase with a weekend of agent work. Combine it with the flat EBIT line and the pattern is individual gains that never reach the income statement. Ask which of your product's SKUs a customer's own engineers could stand up in two weeks, and move your roadmap toward the parts that require data or distribution they cannot assemble.
Confidence: Established fact — a published survey with a stated sample of 1,719 respondents, though self-reported and reliant on respondents' own attribution of EBIT effects.
Go deeper McKinsey · The Register
Technology
Short-lived, task-scoped credentials for agents are now generally available
Vercel moved Connect to general availability, replacing long-lived API tokens for agents with credentials issued at runtime, scoped to an individual task, and expiring automatically. The release added more than 100 connectors, a single integration model, and production governance controls. The problem it targets is credential sprawl: agents accumulating standing access to systems long after the work that justified it finished. It arrives the same week ChatGPT Work started authenticating into third-party sites on a user's behalf.
So what
The pattern here is worth more than the vendor: ephemeral, task-scoped credentials are becoming the expected answer when a customer's security team asks how your agents hold access. Most teams are still handing agents the same long-lived service tokens they gave background jobs five years ago. Have your engineering lead inventory every standing credential currently held by an agent, then set a policy that any new agent integration issues scoped credentials with an expiry, whichever vendor you use.
Confidence: Established fact — the release and its capabilities are documented by Vercel, though the security benefit in practice depends on implementation.
Go deeper Vercel · TLDR AI
