The other side gets a turn 

Last week, Blue Cross Blue Shield attributed ballooning costs not to more or better care, but to hospitals using AI-powered medical coding tools to find more diagnoses to bill for.

A few days later, Apollo’s chief economist suggested that AI assistants like Meta’s Muse could induce an “agentic bank run” by automatically moving household cash out of checking accounts paying almost nothing into higher-yielding accounts.

Investors are already trading on this “humans are lazy, but agents aren’t” thesis. Bloomberg reported last week that shares of banks, insurers, telecom companies, travel platforms (and even Planet Fitness!) fell as Muse took off. Goldman Sachs has assembled a basket of companies exposed to so-called “consumer inertia” that clients can short.

But let’s remember our game theory: the other side gets a turn. Banks can change the terms on their accounts. Insurers can deploy their own agents to scrutinize claims. This isn’t the end, this is just the beginning. So to find interesting opportunities, we need to consider how incumbents might respond. 

I see three possibilities:

The seller changes the offer. If agents insist on moving money around to seek higher yields, banks can just offer their best rates to customers who bring larger balances or commit to leaving them in place longer.

The counterparty enforces the rules. I wrote about this a few weeks ago when Resy began cancelling accounts that used agents to hammer its reservation book. If a platform controls access to scarce supply, it can decide how agents are allowed to request it. Muse’s recent partnership with OpenTable shows where this is heading. With Muse, rather than pinging Resy’s API endlessly, an agent books on your behalf through OpenTable, following rules the platform has defined.

The parties enter an arms race. This is what Blue Cross Blue Shield is warning us of. A hospital’s AI finds a diagnosis that supports a higher bill. An insurer’s AI finds grounds to dispute it. The hospital’s agent appeals. Everyone gets better at arguing, and tokens get burned, but we end up more or less in the same spot as before.

The common thread across all three is that agents make asking free. But asking isn’t transacting. An agent still needs a counterparty that can answer, and in most markets, very little supply is set up to do so seamlessly and programmatically. As demand becomes more fluid and immediate, programmatic supply becomes the scarce thing.

This is the next great opportunity for startups, in my view, and it extends well beyond consumers. The same dynamic applies anywhere one business buys from another. One way to take advantage of this opportunity is to simply create new capacity outright. Another is to solve a hard problem for existing suppliers and, in so doing, make their capacity available to transact with on-demand.

A founder I spoke with recently is building exactly this in the form of a factory scheduling product. Deployed at sufficient scale, this company will know which machines across hundreds of factories can take a job at any given time. A sourcing agent can generate the request, but it needs a counterparty who can reliably say yes, give a price, and deliver. Getting that supply into the market has been the hard part, but for this company, AI is finally making it possible.

In other words, while everyone is building an agent that can ask for anything, perhaps the interesting opportunity is to build the company that can actually say yes?

David Peterson

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

HARD TECH

Google's TPUs head to orbit. Project Suncatcher's first satellite launches October 1 via SpaceX, carrying four TPUs powered by 1 kW of solar panels. Built with Planet, it will test how the chips handle launch loads, radiation and heat, running Gemini workloads in 15 minute intervals to allow for thermal recovery. SpaceX is planning its own orbital compute satellites for 2027.

China's most powerful chip, by Alibaba's count. Alibaba unveiled the Zhenwu V900, claiming 3x the performance of the M890 and clusters of up to 500,000 chips, and set a target of more than 20GW of global cloud capacity by 2032. Chinese hyperscalers are building their own silicon stack partially due to export controls that impact Nvidia. Manufacturing is the open question; Bloomberg reports it is unclear whether Alibaba will work with SMIC.

From smart glasses to data centers. Dutch Morphotonics raised €40M+ to scale nanoimprint lithography, which already produces waveguides for AI glasses. The company is now moving into co-packaged optics (optical interconnects packaged directly with chips). Manufacturing processes built for consumer displays are becoming a route to volume for AI data center interconnect.

HARD MARKETS

The battle over data centers. Data Center Watch counted 45 US data center projects worth $68B blocked or delayed by local opposition in Q2, with 843 opposition groups active across 49 states. Zoning, power and community consent now constrain data center buildouts as much as chips do.

Europe's version: waiting for the grid. Nscale's Loughton site, which the UK government billed as its largest sovereign AI data center, will not get its 90MW of power until the early-to-mid 2030s, against a 2027 launch target. 315 UK data centers are waiting for grid connections and together request 73GW; UK peak demand is 45GW.

Cheaper intelligence. Anthropic released Claude Opus 5.5 at 4/20 per million input/output tokens, 20% below its predecessor. OpenAI's GPT-6 Sol (2/10) and Luna (0.10/0.50) launched at half the price of GPT-5.6, and OpenAI says the prices are permanent. Falling model prices widen margins for application companies; the pressure lands on the labs.

DeepSeek momentum. DeepSeek hit a $1B annualized revenue run rate and is finalizing a $7.5B raise at a 500B-yuan valuation. Its API gross margin was 82.9% in July. The lowest-priced frontier lab is also one of the most profitable on inference, which makes the price war harder for everyone else.

HOW TO STARTUP

Inside an AI-native org. Elena Verna shared field notes from Lovable: granular titles are disappearing, autonomous individual contributors are replacing layers of middle management, and fewer meetings mean faster shipping. It is a practical reference for founders designing lean teams.

Don't throw slop grenades. Shane Parrish proposed a golden rule for AI at work: "Don't use AI to save yourself time by wasting someone else's." Shopify CEO Tobi Lütke named the failure mode: passing on unchecked AI output so a colleague has to clean it up. Teams that adopt AI need a norm that the sender owns the output.

Pricing on outcomes. Ema raised a $77M Series B led by Creaegis for AI agents that do HR, IT and finance work. It charges for completed tasks and results rather than seats or tokens, and reports 180% net dollar retention and 80% gross margins. When agents replace work instead of assisting users, seat counts stop tracking value.

HOW TO VENTURE

Defense gets repriced. Tekever raised a $580M Series D at $6.4B, led by UC Investments and Baillie Gifford, up from a £1B valuation in May 2025. The Lisbon-founded drone maker recently won a British Army surveillance contract worth up to £400M over ten years, and its CEO plans acquisitions in anticipation of consolidation across European defense tech. Late-stage capital is now pricing European defense companies on contracted government demand.

Moonshot is developing an electromagnetic launcher to send smaller, frequent cargo shipments into space.

Groundcover acquires Israeli startup Wand for tens of millions of dollars.

Groundcover
Senior Product Manager (Tel Aviv)

Levity
Forward Deployed Engineer (Germany, Remote)

Motorica
Platform Engineer (Stockholm)