In July 2026, we covered the release of Opus 5, three Gemini models, FLUX 3, Kimi K3, a DeepSeek update, plus a novel mathematical proof and an unprecedented security incident. In one month. This pace is anything but natural: most industries take years to deliver a product generation. Here's why this one is moving so fast.
The question deserves better than a vague answer about technological progress. The acceleration of AI can be explained by precise mechanisms that reinforce each other. Understanding their structure also reveals what could slow them down.
Loop 1: AI accelerates its own construction
This is the most direct loop. Researchers building models use models to work. They write code with agents, analyse results with assistants, explore avenues with reasoning systems.
Each generation makes the next faster to produce. This isn't yet self-improvement in the strong sense, where a system rewrites itself, but it's a very real leverage effect on team productivity. Labs explicitly measure this capability in their safety evaluations.
Loop 2: competition rules out pausing
The market now has enough credible players that no one can afford to slow down. We saw it throughout the month: Chinese labs are publishing open models that rival the best closed ones, forcing Western players to cut prices and ship faster.
This dynamic produces visible effects. A model that launches cheaper than its predecessor is a classic economic anomaly, but it's becoming the norm here. And a delay costs reputation immediately, as Google showed with its flagship delayed three times.
The consequence is structural: even a player who would like to take more time to test properly can't without ceding ground. That's precisely what makes safety debates so difficult.
Capital flows to sectors where progress is visible. That capital funds compute, talent and infrastructure, which accelerates progress further, which attracts more capital. We've documented the scale reached with the financing structures around data centers. This is the most powerful loop in the short term, and also the most fragile: it reverses brutally if confidence turns.
Loop 4: knowledge circulates immediately
This one is more discreet but perhaps the most durable. Unlike the pharmaceutical or aerospace industries, where advances stay locked behind patents for years, a large share of AI research is published within weeks, sometimes days.
An idea found in one lab is picked up, tested and improved elsewhere almost immediately. The publication of weights, as with Kimi K3, amplifies the phenomenon further: it's no longer just the idea that circulates, it's the complete artefact, ready to be dissected.
What could slow things down
Three obstacles stand out, and they're physical rather than theoretical.
Energy and silicon. You can print money, not gigawatts. Building power plants and data centers takes years, and that slowness can't be worked around. This is probably the hardest constraint.
Data. The reserve of quality human text isn't infinite, and the web is filling up with generated content, with the risk of impoverishment we described in our article on model collapse.
Trust. This is the least predictable brake. A major accident, a scandal, or simply disappointment on financial returns could reverse the capital loop within months. Current speed rests on anticipation, and anticipations correct themselves.
What to take away
The acceleration is neither magical nor inevitable. It's the product of four loops feeding each other, three of which could slow down for very concrete reasons.
There's a practical consequence for each of us. In a field moving at this pace, memorising the state of the art makes little sense: what you know about the best model will be outdated in six weeks. What remains useful are the principles: understanding what a token is, why an AI hallucinates, how a real cost is calculated. Those mechanisms don't change at the pace of announcements, and they let you read every new release without starting from scratch.
That's the bet behind what we write here: covering the news, but above all giving the keys that make it legible. Because at the current pace, that's the only thing that holds up.