Eight + One Predictions for AI in 2030

I know we’re famously bad at predicting things, including recessions. So let’s treat these predictions for AI in 2030 as ideas I’ve pieced together from private conversations about where AI is going over the next 3–5 years.

1. AI models will become commodities

We still argue about which lab is ahead: OpenAI versus Anthropic, Opus versus whatever just came out. Then we spend another week arguing about the benchmark.

I expect most of that to disappear from the customer’s view. For ordinary work, several models will clear the required quality bar. The product will choose one, switch when something better or cheaper arrives, and carry on. People using it may never know which model did the work.

I’ve written about this as the new rebar. Models remain useful after competitors reproduce their capabilities. A lab with a three-month lead may struggle to charge a premium for years.

That’s why I’m skeptical that the labs will capture most of the value. Distribution, customer relationships, integration, and the knowledge needed to do the job look more durable. Labs can build those businesses too. Having the best model alone won’t be enough.

2. You will pay for outcomes, not tokens

When I send a video on WhatsApp, I care that my friend gets it. I don’t choose the server or inspect the codec. I expect AI products to become similarly boring underneath.

You’ll pay for a resolved customer problem, a code change that passes review, or an invoice correctly entered into the system.

Cheap subscriptions can hide that calculation for a while. Altman said ChatGPT Pro was losing money in early 2025. Claude already lets paid users buy extra usage at standard API rates. I don’t expect VC investors to keep absorbing the bill for heavy users indefinitely.

Products like Fin and Sierra already price agreed outcomes. With Meta’s strategy, Zuckerberg can earn a return through advertising, better products, and independence from other model providers. He doesn’t have to charge for every token.

3. AI agents will talk to agents

Your next customer may send an agent to find suppliers, request quotes, and compare terms. A human can make the final decision and still never see your company.

Having the best offer won’t help if the buyer’s software can’t discover it or understand it. Businesses will need to make prices, availability, capabilities, and constraints accessible to agents. A2A is an early attempt to make agents communicate across systems.

Customers can generate detailed requests and complaints continuously, at very little cost. German Jobcenter managers have reported lengthy, apparently AI-assisted appeals. Some contain erroneous legal citations.

AI can help people exercise their rights. Writing a request becomes cheap while investigating it still takes effort.

Businesses will need agents that can retrieve evidence, handle routine requests, and return consequential decisions to people.

4. Local and sovereign AI will become the norm

For routine internal questions, a bank needs a capable model it can operate, afford, and keep available. Other industries, like defense and pharma, have their own reasons to care about where data goes and who can turn access off.

As open-weight models become sufficient for more work, more companies will run them on their premises or use private or sovereign hosting. For banks, sovereignty already reaches into exit plans and provider dependence.

5. Managers will manage teams of AI agents

Imagine a customer-service agent with access to order records and permission to approve refunds below €100. It handles the routine cases. A disputed delivery, an ambiguous policy, or a larger refund comes back to its manager with the evidence attached.

Two or three times a day, the manager reviews those cases, decides, and updates the instructions. The manager also assigns work, defines authority, and checks whether the agents did the job properly.

I expect mixed teams to become common. Some tasks go to people, some to agents. The manager needs to understand what each can do and where a failure would matter. Giving an agent access to five systems creates management work before it saves any.

Agents need the decisions and context behind the work, with clear ownership and permissions.

6. Some work will go back to humans

Being able to automate a job doesn’t necessarily mean it’s worth automating. I expect some companies to assign work to agents, count the full cost, and move parts of it back to people.

Take a routine engineering workload. Compare an ordinary engineer with an agent using an expensive model. Count inference, integration, supervision, retries, and correction. Apply the same quality requirement to both. A cheap first attempt can become an expensive accepted result.

In METR’s small early-2025 trial, experienced developers working in familiar repositories took longer with AI. Its later work suggests newer tools help more, though the follow-up was harder to interpret.

Models are also getting cheaper, and efficient routing will favor automation in plenty of cases. But companies will discover workflows where the checking and cleanup consume the apparent savings.

7. The AI industry will consolidate

I don’t expect the current collection of frontier labs and overlapping products to arrive in 2030 in its present form. Some will merge, get acquired, narrow their scope, or disappear.

Labs and infrastructure providers in the capex arms race are making large commitments with uncertain payback. In the announced NVIDIA–OpenAI partnership, NVIDIA proposed investing in OpenAI while OpenAI planned to buy NVIDIA’s infrastructure. I want to know how much demand stands on its own when those financing relationships change.

a16z reported that roughly 2% of US households paid for an AI service in April 2026. Still, I suspect that even universal adoption won’t repay the current buildout. I do think falling model prices make today’s commitments harder to justify.

I doubt both OpenAI and Anthropic keep today’s breadth, ownership, and business model through 2030. Useful AI can grow while the companies financing it change substantially.

8. Generalists will orchestrate; specialists will execute

We won’t need the largest general-purpose model at every step. A strong generalist can plan, deal with ambiguity, and decide what happens next. Smaller systems can do much of the execution.

Sara Hooker’s scaling critique, which I wrote about, questions treating more training compute as the main route to progress. The return also depends on data, algorithms, and architecture.

Jev returns typed judgments. Cloudflare’s Clef makes structured decisions. Microsoft-Decision-1 is in preview for predefined choices and classification. Google takes another route with embedding-based decisions on the device.

A model can route a ticket or check whether an invoice fits a policy without generating an essay. I’d also expect workers optimized for jobs like refactoring Python, with their models chosen for that workload.

A narrowly assigned agent can still use a generalist. But the default of sending every step to the biggest model looks expensive.

+1. Better answers, weaker intellectual communities

Beyond the technical and economic changes, I think AI could make people better informed while making them less likely to contribute to the communities that produce knowledge.

Dan Williams argues that LLMs could move public opinion toward expert knowledge, particularly compared with social feeds. A patient explanation can be more useful than another viral argument, even with the errors and biases these systems still have.

Fernando Borretti argues that intellectual activity depends on a human community. People need material to build on, an audience, recognition, and the sense that their work helps someone.

Someone gets an excellent explanation from an agent and never encounters the person whose tutorial helped make it possible. Fewer readers become contributors. Fewer beginners learn through helping others.

Keynes wrote that “in the long run we are all dead.” In the meantime, I’m looking forward to seeing what we build and which of these guesses we make obsolete.