The AI Moat Has Moved
Why the next decade of AI won’t be won by the company with the smartest model
I’d spent days reading AI announcements: new models, funding rounds, infrastructure deals, benchmarks, acquisitions, enterprise rollouts. Almost none of the announcements that mattered were about making AI smarter.
On June 22, Micron and Anthropic signed a four-part deal: co-designing memory and storage architecture, a multi-year supply agreement for HBM, DRAM, and SSDs, a company-wide rollout of Claude inside Micron, and a Micron investment in Anthropic’s Series H round, the raise that pushed Anthropic’s valuation close to a trillion dollars.
Two days later, Qualcomm paid roughly $3.9 billion for Modular (a compiler company, not a chip company) and opened talks to buy Tenstorrent for as much as $10 billion, assembling a combined $14 billion answer to NVIDIA’s software stack in a single week.
Around the same time, Anthropic disclosed that more than 80% of the code its own engineers merged in May was written by Claude, not by people.
Individually, interesting stories. Collectively, a bigger one. For three years, we asked one question: which model is the smartest? The more important question now is: who owns the layer above the model? That’s where the next wave of advantage will be built.
Every Technology Wave Starts with Capability and Ends with Ecosystems
Cars competed on horsepower. Cloud providers on virtualization. Smartphones on touchscreens and cameras. Then those capabilities became table stakes. Customers stopped buying the product with one more feature and started buying the one that fit their lives: the one with the ecosystem, the integrations, the reliability. The thing that was hardest to rip out. The winners were rarely the companies with the best technology. They were the companies that became the hardest to replace.
AI is reaching that moment far faster than most people expected, because models are improving faster than enterprises can change. In the first quarter of 2026 alone, the major labs shipped roughly 255 models, about three a day. Any workflow hard-wired to a single model is taking on technical debt by the week. Companies don’t rebuild their operations every six weeks. So the workflow stays, the model swaps out underneath it, and the value quietly migrates from the intelligence to the thing that survives every upgrade.
The Moat is Moving in Four Places
1. Infrastructure: Buying Visibility Over Silicon
Micron’s deal with Anthropic looks like a chip company landing another customer. It’s really a chip company buying visibility. Frontier models increasingly live or die on memory bandwidth and architecture, not just GPUs, and Micron’s high-bandwidth memory is already sold out.
By co-designing the memory subsystem for how Claude actually trains and runs, Micron sees where AI is heading long before the rest of the industry does. It’s telling that all three of the world’s HBM makers (Samsung, SK Hynix, and now Micron) are Anthropic investors. The asset isn’t better memory; it’s being closest to production. Whoever learns fastest eventually wins.
2. The Developer Ecosystem: Selling an Escape Hatch
Qualcomm’s aggressive hardware plays get filed under silicon, but neither Modular nor Tenstorrent is really about chips. NVIDIA’s moat was never the GPU; it was CUDA, which has locked in roughly four million developers for the better part of two decades.
Modular is the compiler company founded by Chris Lattner (creator of Swift and LLVM), and its MAX stack lets developers write AI code once and run it across NVIDIA, AMD, Intel, and Apple silicon without rewriting it. Tenstorrent designs inference chips on the open RISC-V standard. Put the two together and you get what every previous NVIDIA challenger lacked: not just a faster chip, but a way out of the ecosystem. Developers don’t buy chips, they buy productivity, and habits are the hardest thing in computing to change. Qualcomm is buying an escape hatch from someone else’s ecosystem, because ecosystems outlast every individual technology lead.
3. Workflow: Execution is the New Unit of Value
This is the biggest shift of all, and it’s where the whole argument lands:
The application is becoming the product. The model is becoming infrastructure.
The strongest evidence came from Anthropic itself. More than 80% of the code its engineers merged in May was written by Claude, roughly eight times as much code shipped per engineer as a few years earlier, and effectively 100% on some products. Anthropic’s own chief product officer, Mike Krieger, described it plainly: Claude is now writing Claude.
It’s already happening. Support organizations are moving from assisted replies to autonomous resolution. Banks and insurers are automating entire multi-step processes rather than isolated tasks. McKinsey finds roughly a quarter of organizations are already scaling agentic systems that plan and execute work, and by mid-2026 about a third were running at least one agent in production. Anthropic’s own coding agent, Claude Code, has grown into a business generating billions in annualized revenue. Proof that execution, not assistance, is the product people now pay for.
Companies aren’t buying intelligence anymore. They’re buying execution. That’s an entirely different market.
4. Orchestration: Model Selection as Air-Traffic Control
Almost no enterprise is standardizing on a single model. One writes code, another summarizes meetings, another handles customers, another drafts documentation, each workload routed to whichever model wins on quality, latency, governance, and cost. More than a third of enterprises now run five or more models in production, treating model selection like air-traffic control: send most traffic to a cheap, fast model, reserve the frontier for the hard cases, and you can match single-frontier-model quality at a fraction of the cost.
Cloud went the exact same way: few applications depend on one provider, and users never notice the difference. Model routing is becoming an optimization problem. The application remains the durable asset.
What Should Product Managers Build?
A year ago, I evaluated AI products by asking: Does it use GPT? Is Claude better than Gemini? Which benchmark does it lead? Now I ask one question instead: If OpenAI, Anthropic, Google, and DeepSeek all became equally capable next year, would customers still choose this product? If the answer is no, you own a feature. If the answer is yes, you may own a moat. In a year when no single model dominates every task and specialization is the norm, the questions worth asking aren’t about model quality. They’re about product durability:
Workflow Gravity: Are we embedded in the customer’s daily execution loop?
Context Flywheels: Are we accumulating proprietary data that makes every interaction better?
Process Friction: Are we removing friction across an entire business workflow, or just optimizing one isolated task?
Switching Costs: Are we building integrations, governance, and reliability that survive the next model upgrade?
Five years from now, almost no executive will remember which model topped a benchmark in 2026. They’ll remember which companies became impossible to remove from their work. The first chapter of AI rewarded the companies that built remarkable intelligence. The next will reward the ones that make that intelligence invisible, reliable, deeply integrated, and indispensable.