Forecasting AI is a humbling pastime, but a year out is close enough to be useful and far enough to be wrong in public. Here's my honest call for the twelve months to November 2027 — what's near-certain, what's likely, what's a coin flip, and what I'd ignore. Discount everything appropriately.
Near-certain: the agent plumbing matures
Agents move from impressive demos to boring infrastructure. Scoped-token authorization for agent spending, better sandboxing, standardized tool protocols and real audit tooling all get good enough that mid-size companies ship agents without heroics. Expect a wave of agent-security incidents anyway, because production teaches what labs can't. Frameworks consolidate; half the orchestration libraries from 2025 quietly lose. Winners make evals and replay first-class.
Near-certain: regulation gets enforced, not just debated
EU AI Act obligations keep phasing in, and the first genuinely consequential enforcement actions land — likely around transparency failures or high-risk systems without proper oversight, not sci-fi scenarios. In the US, expect more state laws, more agency guidance and at least one headline-grabbing settlement over AI claims. Compliance theatre will grow alongside the real work. The cost of doing AI properly on paper rises; the cost of doing it well barely changes.
Likely: the price floor drops again
Inference costs per token keep falling as new silicon scales and efficiency stacks up. Open-weight quality keeps compressing the mid-tier: models that were flagship-grade eighteen months ago are table stakes. Expect proprietary labs to defend the frontier with usage-based bundles, enterprise agreements and agent platforms rather than raw token pricing. The bargain segment becomes genuinely excellent — the previous generation of open models at distilled prices will be very good indeed.
Likely: the open/closed frontier stays close, with a catch
By late 2027, the best open models remain within striking distance of proprietary frontiers on most benchmarks, though the proprietary lead persists on the longest-horizon agentic tasks and the wildest multimodal work. The real split becomes provenance: who trained on what, with which licenses cleared. Expect serious releases with unusually detailed data documentation as legal pressure mounts, and expect at least one big lab to be publicly embarrassed by training-data provenance. Hedge on any claim that the gap will “close completely" — both sides keep moving.
Likely: consolidation, and one or two casualties
The market can't support every 2024-vintage AI startup. Expect acqui-hires, distressed sales and a couple of shutdowns among model labs that raised on vibes. Meanwhile, the hyperscalers and a couple of well-fed frontier labs keep outspending everyone. Nvidia stays dominant but fights harder on margins; its custom-silicon rivals win more inference share inside clouds. Power and grid capacity, not chips, gate the biggest data center builds.
A sober look at capability plateaus
Capability gains have noticeably decelerated on pure pretraining scale, and the action moved to post-training: reasoning traces, reinforcement learning from verifiable outcomes, tool-use training and agentic rollouts. That's where the meaningful jumps came from this year, and where they'll keep coming from. If someone claims next year's model is 10x smarter because of a bigger run, check the recipe. Expect steady, compounding gains — better reasoning per dollar, longer reliable horizons, fewer catastrophic hallucinations in long chains — rather than a vertical leap. Which is actually the healthier trend: incremental reliability beats sporadic moonshots for real software.
The talent and org-chart shuffle
Headcount plans across tech keep bending around AI productivity claims, half of which hold up and half of which don't. The durable pattern: fewer people on rote implementation, more on spec, eval and review. Junior pipelines are the worry — if nobody writes glue code for two years, who becomes a senior engineer? The good teams are pairing AI with deliberate apprenticeship rather than pretending the junior rung can vanish. Expect that conversation to get louder next year, especially as budget season opens.
Also: energy politics. AI datacenters are becoming a ballot-box issue in the US and Europe, with interconnection queues stretching into the 2030s in some grids. Expect lawsuits, moratoria in specific counties, and genuinely interesting engineering around flexible, location-aware workloads — training that only runs when the grid is quiet, inference pinned to underutilized regions. It won't stop the buildout; it will map it.
What my gut says actually breaks
Predictions are cheap, so let me add the failure modes nobody wants to own. Something big and embarrassing hits the agent category: a widely deployed agent books, buys or sends something it categorically shouldn't, and the story runs for a week. A frontier lab misses a promised capability date and the market reprices the whole sector for a month. An open-weights lab gets hit with a provenance lawsuit that reshapes licensing. And somewhere in the year, someone ships an AI feature with real users that quietly fails in a way that was obvious in hindsight and invisible beforehand. None of these stop the trend; all of them bite teams that skipped evals and oversight because the demo looked great.
What a model release cycle looks like this year
Release cadence has compressed to where tracking every model is a full-time job. A realistic month might include: a flagship proprietary refresh, two open-weight drops of different sizes, a coding-specialist fine-tune, a reasoning model with a new training trick, and four benchmark claims that all contradict each other. Practitioners have stopped reacting to launches. The winning habit is a quarterly re-baseline: pin your models for a quarter, eval the new candidates against your golden set, swap only what demonstrably wins, and keep the old config ready to roll back. Your users don't care about your model version; they care that Tuesday's answer isn't worse than Monday's.
The capex question hangs over everything
AI infrastructure spend keeps setting records, and the financing is unusually circular — chip vendors, cloud credits and lab equity all flowing into each other's future revenue. It probably works out; it might not. The tell will be whether AI revenue at the application layer keeps compounding at the rate the infrastructure layer assumes. If it does, the buildout is vindicated and prices keep falling on competition. If it doesn't, expect a shakeout where the survivors are whoever locked in capacity and contracts early. For builders, the actionable hedge is the same as always: avoid single-vendor dependency, keep unit economics visible, and don't scale headcount on usage assumptions alone.
Multimodal gets boringly useful
The multimodal frontier stops being a party trick and becomes plumbing. Document understanding that actually reads the ugly scan, voice agents with acceptable latency, and video-adjacent analysis for security and ops workflows all cross the "deploy it" threshold this year. Open models are closing fastest here, which matters because multimodal workloads are brutally expensive per token on proprietary APIs. Expect a lot of unglamorous value: invoices read correctly, field photos checked against specs, support calls summarized accurately. Nobody puts that in a launch keynote; everyone puts it in the ledger.
Coin flips worth watching
- A major lab ships a genuinely reliable, general computer-use agent product to consumers — or has to pull one back after a public failure.
- A US federal AI bill advances meaningfully, or the patchwork hardens and preemption fights get louder.
- A notable open-weights release matches the proprietary frontier across the board for a month, however briefly.
- Energy and interconnection become the explicit public villain of AI economics, with real political consequences.
What I'd ignore
AGI-by-Christmas announcements. Benchmark records trumpeted without task relevance. Any forecast (including this one) phrased with total certainty. And the idea that you must rebuild your stack every quarter — the durable wins come from teams that integrate carefully, evaluate honestly and ship steadily while the hype cycles churn.
Two final calibration notes. First, I've been deliberately specific about what's near-certain versus speculative, and I want to keep those categories clean: enforcement, consolidation and cheaper tokens are where I'd put my name on the table; AGI timelines are where I'd ask you not to take anyone's word for it, including mine. Second, the meta-pattern of the last three years is that AI progress is lumpy in public and smooth in private — demos leap, deployment grinds. Twelve months from now the benchmarks will say one thing and production logs will say another. Build for the logs.
If you're placing one bet only
Forced to pick a single theme for the next twelve months, I'd bet on agent reliability infrastructure — evals, replay, guardrails and metering — being the most underpriced layer in the stack. Capability is nearly commoditized; dependable execution is not. Whoever sells boring trust for autonomous software eats well while the model wars take the headlines. Everything else on this page is commentary; that's the thesis.
Pricing power migrates to distribution
One meta-prediction worth flagging: as model capability commoditizes, the durable moats shift away from the weights and toward distribution and workflow ownership. Whoever already owns the user surface — the IDE, the CRM, the helpdesk, the browser — gets to decide which model answers, and takes the margin. Expect the labs to respond with deeper bundling: agents sold inside productivity suites, inference wrapped in enterprise agreements, and usage credits that quietly lock in loyalty. The independent builder's counter is focus plus trust — own one workflow so completely that switching costs stay high even when the underlying model is swappable.
It's worth watching where developer mindshare actually accrues this year: not the launch-day leaderboard, but the tools teams rewire their CI and support queues around. Distribution wins are boring and sticky; benchmark wins are loud and perishable. Plan your bets accordingly. And keep an eye on the open ecosystem here — the projects that become default infrastructure (package registries, model routers, eval runners) will quietly capture more durable value than most model launches, the same way npm and PyPI became load-bearing without ever trending. The compounding advantage accrues to whoever controls the last mile between the model and the user's actual job — and that last mile is rarely the model itself.
The through-line
The next twelve months belong to whoever turns capability into reliability. Cheaper tokens are coming regardless; the differentiator is trustworthy deployment — agents that don't surprise you, compliance you can show, and costs you can predict. If 2025 was the year of the model, 2027 is the year of the harness. Build accordingly.