Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional deep dive that turns out to be sharper than the analysis around it. Here’s what stood out between 22 and 29 September, and why I think it matters if you are running agentic AI in production.
The shape of it: the median American company spends $12.50 per employee per month on AI, and still cannot find the return. Meta stopped selling an agent this week and started selling the stack behind it. Shopify opened its checkout to browser agents while Amazon kept its doors bolted. Last week enforcement arrived as somebody else’s default setting. This week the bill arrived — in permits, in interest rates, and in the gap between what AI costs and what it returns.
Meta stopped selling an agent and started selling the stack
TechCrunch’s Monday evening brief on 28 September carried the story that changes my week: Meta launched an enterprise AI platform and hired MongoDB’s CEO to lead the new initiative. Meta says it will focus on bringing its full technology stack to businesses and developers — Muse, Meta Business Agent, Muse API and Muse Code among them.
Read that against the consumer numbers Chamath Palihapitiya assembled in his 27 September note. Muse launched on 8 September, runs on its own computer in Meta’s cloud and keeps working after you close the app. Ask it for a product and it finds one and checks out, logging in as you. It reached number one on the US App Store and has passed 3.4 million downloads. At Connect, Meta added Walmart, Best Buy, Sephora and Wayfair as partners. It has a free tier and paid plans at $20 and $100 a month, and Chamath quotes Zuckerberg saying Meta will keep it free for “a huge number of tokens,” expecting that “over time we will profit by taking a small fee from transactions.”
Three weeks from consumer launch to enterprise platform is not a product roadmap, it is a land grab — and the hire tells you what they think they are selling. You do not recruit the chief executive of a database company to sell a chatbot. You recruit one to sell the layer underneath. Which is precisely the layer I keep arguing you should own. The pitch will be seductive: one API key, one business agent, one stack, your data in their harness. Ask the question that survives the demo — if Meta reprices the transaction fee, deprecates Muse Code, or decides your category competes with a partner, what of yours moves and what of yours stays? If the honest answer is “all of it moves,” you have not bought a platform. You have rented a business model.
Amazon bolts the door, Shopify opens the till
The clearest split of the week is in commerce. Chamath sets it out with the numbers attached: Amazon made $19.8 billion from advertising in the second quarter, including the sponsored listings shoppers see as they browse, and on 20 September it began blocking Muse from its store. Shopify went the other way. It earns when a sale happens, wherever the shopper started, so it plugged its checkout into Muse — and Shop Pay is being added at checkout alongside the Stripe payments Muse launched with.
Then, on 28 September, TechCrunch reported the general version of the same decision: Shopify is expanding WebMCP support to checkout, allowing browser-based AI agents to update order details and complete purchases with a buyer’s authorisation. Not one partner agent. Any agent that speaks the protocol.
The market has already priced the split. Chamath notes that since Muse launched, Meta’s stock is up about 22%, while TripAdvisor, Bookings, Planet Fitness and The New York Times fell. Expedia and Instacart are down despite joining as partners. Shopify, which runs the checkout, rose 11.4% in the week of Connect.
Chamath’s explanation is the one I would put in front of a board: agents bring price discovery and transparency, and that is bad for anyone who profits from opacity and breakage — the subscription that is impossible to cancel, the fee you only find at step four. So the block is not about safety. It is about margin. The lesson for the rest of us is structural, not moral. Every surface your agents touch is now choosing a side: own the customer’s screen, or get paid when the transaction clears. Firms in the first camp will block you. Firms in the second will publish a protocol and invite you in. Work out which camp each of your dependencies is in before your roadmap assumes the wrong one. And notice which side is publishing an open protocol — WebMCP is doing more for agent commerce this quarter than any model release.
The model layer repriced again, and nobody flinched
Three items from the same fortnight, which belong together. Anthropic released Sonnet 5.5, the newest version of its midrange model, which it calls a significantly cheaper, faster work partner — TechCrunch’s summary is faster response times and less token burn. Google is killing off Gemini’s Gems in favour of “skills”, and TechCrunch gives the reason plainly: as all-in-one AI agents like Meta’s Muse and Instinct take off, Google is ending the feature that built task-specific agents. And AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion, with Li joining AMD as executive vice president and chief scientist.
Cheaper midrange. The deprecation of a whole way of configuring agents. A world-model company absorbed into a chip company. All in a week that most people will remember for a shopping agent.
The Gems story is the one to sit with, and almost nobody will. Google just deprecated a container that thousands of teams had already filled with their own instructions, examples and scope rules. That work was context — the expensive, hard-won kind — and it lived in a vendor’s product surface rather than in a repository the customer controls. Whatever the migration path to “skills” turns out to be, everyone who built there is now doing unplanned work because a product manager somewhere changed a strategy. This is the whole argument in one news item. Rent the model. Rent the compute. But if your agent’s instructions, tools and guardrails only exist inside a vendor’s console, you do not own your harness — you are a tenant, and the notice period is a blog post. Keep the definitions in your own repo, in plain text, under review, and let the runtime be the thing you swap.
OpenAI published its misalignment reports, and the breadth is the story
On Friday 25 September, OpenAI published a new site devoted to “misalignment reports.” TechCrunch’s read on it was not gentle: the breadth of the incidents is alarming, and the headline the brief chose was that OpenAI still doesn’t seem to have a handle on all of its rogue AI activity.
I want to give credit where it belongs. Publishing a standing register of your own model’s misbehaviour is a genuinely uncomfortable thing for a commercial lab to do, and the field is better for it. But a register is not a control. It is a record that something already happened.
Here is the practical translation, and it has nothing to do with whether you use OpenAI. The frontier lab with the most instrumentation, the largest safety team and every incentive to look competent is publishing a list of things its models did that it did not intend. If that is the state of the art at the model layer, then no amount of vendor assurance substitutes for controls at your layer. The uncomfortable question for most enterprise teams is not “is the model safe?” It is: if our agent did something we did not intend last Tuesday, would we know — and could we produce the trail? Most teams I meet could answer the first within a week and the second not at all. A misalignment register you can only read about somebody else’s model is interesting. One you can run against your own deployment is an asset.
The ROI problem is not the token bill. It is the calendar.
Chamath’s 25 September deep dive is the most useful thing I read all week, and the arithmetic deserves to be repeated at every steering committee. Using Ramp’s AI Index, which tracks what more than 70,000 US businesses spend on AI: the median company spends $12.50 per employee a month. Against an average employee cost of about $8,500 a month, AI only has to make that person roughly 0.15% more productive to pay for itself — about three additional productive minutes a week.
So why is the return so hard to see? Because, as he puts it, most of a corporate employee’s week goes to coordination, meetings, alignment and process sign-offs, not to producing work. AI speeds up production. It cannot decide what happens to the time it returns. “AI can hand back those three minutes, but it can’t decide what happens to them.” If they flow into another meeting or another week waiting on legal, the return is zero.
The evidence he assembles for the gap is striking. At P&G, one person using AI produced product proposals as good as a two-person team without it. MIT economists found that automation pays when AI steps are grouped so a person checks the work once, not after every step. And a 2026 experiment across 515 startups gave every firm the same AI tools, training and support; a random half also saw how other firms reorganised their work around AI. For the typical startup, revenue barely moved in either group. Almost all of the gains came from the top 10% of firms, and the treatment group pulled far enough ahead to reach 1.9x the control group’s revenue overall. Identical tools. The difference was how the work was rebuilt around them.
He reaches for two old quotes to close it. Bill Gates in 1995: automation applied to an efficient operation magnifies the efficiency, and applied to an inefficient one magnifies the inefficiency. And Robert Solow in 1987: “You can see the computer age everywhere but in the productivity statistics” — after which US productivity growth went from about 1.5% a year between 1973 and 1995 to about 3.3% a year from 1995 to 2003, once work was reorganised around the machines.
This is the single best argument I have seen for the position I have been taking all year, and it arrives from an investor rather than an engineer. If the token bill is a rounding error and the returns still do not show, the bottleneck was never the model. It is the harness, the process and the record — which is to say, the parts you own. Elon Musk’s five steps, which Chamath quotes, are the right order and almost everyone runs them backwards: question every requirement, delete what you can, simplify what is left, speed it up, and only then automate. Musk says doing it in reverse and automating unnecessary processes was one of his biggest mistakes at Tesla. Most “AI transformation” programmes I am shown start at step five. An agent pointed at a process nobody has questioned is a very fast way to do the wrong work. And the 515-startup result should end the debate about whether to run a discovery phase: the tools were free to everyone, and only the firms that redesigned the work got paid.
The constraint on your agent roadmap is a pipeline in New Mexico
Two of my inboxes carried the same story from different ends. TechCrunch on 25 September: Oracle sent a force majeure notice on its New Mexico Stargate data centre, which would let it delay payments should the facility miss its 2028 target to come online. Chamath, citing Bloomberg’s 24 September report, supplies the detail. The campus, Project Jupiter, is 2.45 gigawatts, designed to run on gas-powered fuel cells from Bloom Energy fed by a 17-mile pipeline. New Mexico’s State Land Office refused to let 0.6 miles of that pipeline cross state land, calling the burden on the state’s water and communities “extreme.” The pipeline is about six months late, the air permit is still pending, and the $18 billion construction loan behind the project now trades below 90 cents on the dollar. Oracle says notices like this preserve its contractual rights and that the project remains on schedule.
The same week, both large states tightened. On 21 September, Texas Governor Greg Abbott ordered the state’s environmental regulator to stop issuing data centre permits until audits of the grid and water supply are complete, with conditions requiring data centres to pay all their own electrical infrastructure costs and to lower residents’ bills. That same afternoon, California Governor Gavin Newsom signed seven bills requiring data centres to report energy and water use, pay for their own grid upgrades, and cover a larger share of wildfire costs.
And underneath all of it, the price of money moved. Chamath notes the US government now pays about 5.18% to borrow for ten years, the most since 2007, and 5.5% for thirty years, the most since 2004. The average 30-year mortgage is 7.03%. The Fed raised rates on 16 September for the first time in three years, the 23 September five-year sale drew its weakest demand since 2018, and the ten-year real yield rose from 2.68% on 18 September to 2.85% on 24 September while expected inflation stayed near 2.35%.
Half a mile of pipeline is holding up 2.45 gigawatts. That is the whole lesson. The agentic AI stack now bottoms out in land, water, permits and the cost of borrowing — four things no lab controls and no model release improves. Chamath’s line is the one to keep: energised land with a signed interconnection is the one input in this buildout that cannot be manufactured on demand. For a buyer this is oddly clarifying. If compute gets scarcer and dearer while capital gets more expensive, the inference-heavy architecture you designed on 2025 assumptions gets a worse bill every quarter, and the vendor absorbing that cost today will not absorb it forever. Design now for the world where compute is expensive: cache aggressively, route the cheap work to cheap models, and make sure changing that routing is a config change rather than a quarter of engineering. That flexibility is worth more than any single model’s benchmark score.
Quick hits
- Instinct raised a $1 billion Series C at a $10 billion valuation, with founder Noah Shinn saying “we’re just getting started.” Chamath’s note explains why the number is that big: Instinct’s agents can talk directly to each other, coordinating things like group trips across everyone’s calendars and arrival times. It raised $250 million at $2.5 billion in August. Four times the valuation in roughly six weeks.
- Ema raised $77 million as AI starts eating into enterprise software and services — $140 million to date, with more than 50 enterprise customers including Google and Microsoft. The interesting part is the category framing: not “AI tool” but a replacement for software and services budget.
- Databricks bought cloud spreadsheet startup Row Zero and says it is scouting for more startups to acquire, adding to a 2026 shopping spree. The data platforms are buying the interfaces where decisions actually get made.
- Modulate raised $25 million for voice models that detect deepfakes, fraud and scams. As agents start making calls on your behalf, verifying who is on the other end stops being a niche security product and starts being infrastructure.
- Greece’s prime minister Kyriakos Mitsotakis told TechCrunch “we’re already fighting yesterday’s battle,” admitting no government is ready for what AI is about to do. Rare candour, and a useful reminder that the regulatory picture is being written by people who say they are behind.
The theme of the week
Last week enforcement arrived as somebody else’s default setting. This week the bill arrived, and it came in three currencies. Chamath priced the returns: $12.50 a month per employee, against work that has not been redesigned, produces nothing you can find in the numbers. New Mexico priced the inputs: half a mile of pipeline against 2.45 gigawatts. And the market priced the position: Meta up 22%, Shopify up 11.4%, and everyone who sat between the agent and the transaction down.
Put those together and you get a single instruction. The parts of your stack that are getting cheaper are not the parts that decide whether this works. Models got cheaper this week. Compute got harder and more expensive. Access got more political. And the return still turned entirely on whether anyone had redesigned the work. Rent the model — it is a commodity and it repriced again on Monday. Own the harness, the evaluations, the controls and the record of truth. This week I would add one word to that: own the process, because Musk’s fifth step is the only one most programmes actually execute, and it is the only one that cannot save a bad first four.
What this means if you’re deploying agentic AI
- Run Musk’s five steps before you run a pilot. Question every requirement, delete what you can, simplify, speed up — and only then automate. If your programme starts at automation, you are buying a faster version of a process nobody defended.
- Get your agent definitions out of vendor consoles and into your repository. Gemini’s Gems were deprecated this week. Instructions, tools, scope rules and evaluations belong in plain text under change control, where a vendor’s product decision cannot delete them.
- Classify every commerce or data surface your agents touch as “owns the screen” or “gets paid on the transaction.” The first camp will block you eventually, whatever they say today. The second will publish a protocol. Plan the fallbacks accordingly.
- Measure the minutes you are giving back, and decide in advance where they go. Time returned into another meeting is a cost, not a saving. If no one owns the redesign of the freed-up hour, cancel the business case.
- Batch your agent steps and check once, not constantly. The MIT finding is a design instruction: group the AI work so a human reviews a completed unit, rather than approving every hop and destroying the economics of the whole chain.
- Build a misalignment register for your own deployment. If the best-resourced lab in the world is publishing a list of things its models did unintentionally, your assurance cannot be a vendor’s. Log it, review it, and be able to produce the trail.
- Assume compute gets dearer, not cheaper. Make model routing and caching a configuration decision. The permit queue in Texas and a loan trading below 90 cents are not your problem this quarter, but they are your unit economics next year.
Frequently asked questions
What did Meta announce for enterprises in September 2026?
TechCrunch reported on 28 September 2026 that Meta launched an enterprise AI platform and hired MongoDB’s chief executive to lead the new initiative. Meta says it will focus on bringing its full technology stack to businesses and developers, including Muse, Meta Business Agent, the Muse API and Muse Code. It follows the consumer launch of Muse on 8 September, which Chamath Palihapitiya reports reached number one on the US App Store and has passed 3.4 million downloads, with a free tier and paid plans at $20 and $100 a month. For enterprise buyers the significant question is not capability but dependency: which parts of your context, tooling and controls would have to move if the platform’s commercial terms or product lineup changed.
How much do companies actually spend on AI per employee, and why is the return hard to see?
In his 25 September 2026 deep dive, Chamath Palihapitiya cites Ramp’s AI Index, which tracks AI spending across more than 70,000 US businesses, and reports that the median company spends $12.50 per employee a month. Against an average employee cost of roughly $8,500 a month, AI only needs to make that person about 0.15% more productive to pay for itself, which is around three additional productive minutes a week. The return is hard to see because most of a corporate employee’s week goes to coordination, meetings and sign-offs rather than production, so time returned by AI is often reabsorbed rather than converted into output. A 2026 experiment across 515 startups found that identical AI tools produced almost all of their gains in the top 10% of firms, with the group that also saw how others reorganised their work reaching 1.9x the control group’s revenue.
Why did Oracle send a force majeure notice on its New Mexico data centre?
TechCrunch reported on 25 September 2026 that Oracle sent a force majeure notice on its New Mexico Stargate data centre, which would allow it to delay payments if the facility misses its 2028 target to come online. Citing Bloomberg’s 24 September report, Chamath Palihapitiya explains that the 2.45-gigawatt campus, Project Jupiter, is designed to run on gas-powered fuel cells from Bloom Energy fed by a 17-mile pipeline, and that New Mexico’s State Land Office refused to let 0.6 miles of that pipeline cross state land, calling the burden on the state’s water and communities “extreme.” The pipeline is roughly six months late, an air permit is still pending, and the $18 billion construction loan behind the project trades below 90 cents on the dollar. Oracle says such notices preserve its contractual rights and that the project remains on its planned schedule.
From Veehive Labs
Seven weeks in, the compressed version still holds: rent the model and the compute; own the context, the controls, the evaluation and the record of truth. What this week added is the word in front of all of them — process. The 515-startup result and Musk’s five steps say the same thing from opposite ends of the market: identical tools, and only the firms that rebuilt the work saw a return. A harness without a redesigned process is just a faster route to the same answer. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.
If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, permissions model, cost ceilings, risks and delivery architecture before the build begins.
- Read the deep dive: Own the harness, rent the model
- Read the last edition: Amazon blocks Meta’s Muse, Cloudflare splits the crawlers
- Explore Veehive Labs enterprise AI services · See our solutions
Build Your Organisation’s AI Capability
Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.
Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.