Bloom AI — sharp takes on the AI era, written in real time.

2026 is the year of agents — and the year most of them break in production

Every major frontier lab — OpenAI, Anthropic, Google DeepMind, xAI, Meta, Mistral, DeepSeek — is shipping some flavor of an agentic stack. The demos are spectacular. The production deployments are not. The gap between "Claude can book your flight" on stage and "Claude consistently books the right flight under your team's compliance policy" is roughly two years of engineering, eval scaffolding, fallback logic, and exception handling that no keynote shows.

The companies that win the agent layer won't be the ones with the cleverest prompts. They'll be the ones with the most boring infrastructure: rigorous internal evals, structured tool-use schemas, deterministic guardrails wrapped around stochastic reasoning, and a customer support process that catches the failures the model misses. The hype is on the model side; the moat is on the orchestration side.

Open weights ate the developer ecosystem; closed models still own the enterprise

Llama, Mistral, Qwen, DeepSeek, and the long tail of open-weight releases have become the default for indie developers, AI hobbyists, research labs, and any team that needs to control inference cost or data residency. Hugging Face is now the npm of machine learning. The bottom of the market has been thoroughly commoditized — what used to require an OpenAI API call now runs on a single GPU under your desk.

At the top of the market, however, the procurement story still favors closed providers. Enterprise buyers want SLAs, indemnification, SOC 2 reports, data processing agreements, and a sales engineer who picks up the phone. OpenAI and Anthropic sell that. The open-weight ecosystem mostly doesn't — and the gap, more than any benchmark, is what's keeping the closed labs' revenue lines vertical.

Evals are the next frontier — and most public benchmarks are already cooked

MMLU is saturated. HumanEval is saturated. GPQA is on its way. SWE-bench is the current darling, with ARC-AGI sitting next to it as the resident "but can it really reason?" challenge. The honest read is that public benchmarks measure how well labs can train against benchmarks. Real product evaluation looks nothing like this: it's domain-specific test suites, adversarial probes built from real user logs, regression catchers that fire when a model update silently degrades your most important workflow.

If you're shipping an AI feature in 2026 without an internal eval harness, your product strategy is "hope." Hope is not a strategy.

The compute story is a geopolitical story now

Frontier model training has crossed the threshold where the bottleneck is no longer algorithms or talent but raw access to GPUs, the energy to run them, and the data centers to house them. NVIDIA's market cap, TSMC's lead times, US export controls on advanced semiconductors to China, the Middle East's emerging role as compute financier, and the build-out of multi-gigawatt training clusters across Texas, Wyoming, and the Gulf — these aren't tech stories. They're industrial policy stories with AI labs as the latest beneficiaries.

Anyone trying to forecast where capability lives in five years should be reading earnings calls from TSMC, ASML, and the major hyperscalers — not just papers from OpenAI and DeepMind.

Taste is the last defensible skill

The cost of producing competent output — competent writing, competent code, competent illustration, competent video — has fallen toward zero. What hasn't fallen is the cost of knowing what's worth producing in the first place. The bottleneck has migrated from production to judgment: which problem to solve, which framing to use, which examples to lead with, which sentence to cut, which detail makes the whole thing land.

This is good news for anyone with strong taste and bad news for anyone who confused output volume with skill. The AI era rewards the person who can say "no, not that one, this one" — and the volume of "not that one" candidates a model can generate is effectively infinite.