The S&P 500 Has an AI Concentration Problem, and Open Source Is About to Make It Worse

Roughly thirty-five cents of every dollar in the S&P 500 sits inside seven companies. Almost half of next year's projected earnings growth comes from the same seven. If you bought an index fund in 2020 expecting diversified exposure to the American economy, what you actually own now is a leveraged bet on a handful of AI labs and their hardware supplier.

That bet rested on a simple assumption: the frontier model layer is defensible. A few companies will keep producing the best AI, charge premium prices for it, and the rest of the economy will rent capability from them indefinitely.

That assumption is breaking right now, in public, and most index investors haven't noticed.

What the moat actually was

The premium pricing on frontier AI was never really about the math. It was about the gap. As long as the best closed model was meaningfully better than anything you could download for free, enterprise buyers would pay five dollars per million tokens instead of one. Engineering teams would standardize on the leader. CFOs would sign the checks.

A year ago, the gap was real. Open-weight models trailed the frontier by something like a year on hard reasoning tasks and were treated as fine for hobby projects but not for production. That's the world the current valuations were priced in.

That world ended sometime in the last six months.

Where things stand now

A short, uncomfortable list:

DeepSeek V4, released earlier this year, beats GPT-4.1 on multiple benchmarks and costs roughly a third of what the equivalent closed model charges per token. Qwen 3.6 from Alibaba runs competitively at 22B active parameters. Llama 4 spans from edge-device sizes to frontier-competitive. GLM-5.1, Kimi K2.6, Mistral Medium 3.5, five frontier-class open-weight models shipped in a single thirty-day window this spring.

Eight of the top ten Chinese AI models are now open-weight. Qwen has overtaken Llama in cumulative downloads. Roughly a third of all open-model downloads globally come from Chinese labs.

Switching costs, which everyone assumed would protect incumbents, turn out to be almost nonexistent. Ramp's enterprise spend data shows fifty-two percent of companies using a frontier provider are using both Anthropic and OpenAI. Forty-three percent of Anthropic's customers switched from another vendor. The narrative of vendor lock-in was, mostly, a story sellers told themselves.

And the model providers know it. The recent wave of consulting acquisitions, partner network announcements, and infrastructure deals from the major labs isn't aggressive expansion. It's diversification. They are racing to become something other than a pure model provider before the model business looks like the airline business.

The misread on chips

The standard counter to all of this is: doesn't matter, NVIDIA still wins.

That's half true. The chips layer is more defensible than the model layer, and NVIDIA has played the open-source wave brilliantly, releasing its own Nemotron family, partnering with the inference clouds that serve everyone's open models, taking its margin every time a token gets generated regardless of whose weights are running.

But the chips story has two cracks. First, training and inference are different workloads. CUDA's moat is strongest in training, and training is the smaller and more concentrated workload. Inference, which is where the actual revenue gets generated, is increasingly running on AMD, on custom silicon from Google, on dedicated inference startups, on anything that can serve tokens cheaply. Inference-layer compilers and runtimes are explicitly designed to be hardware-agnostic. Second, NVIDIA currently trades at a multiple that assumes ninety percent share forever. Eighty-five percent share at lower per-chip margins is still a great business, but it's a different stock.

What this means for the index

The labs don't have to die. NVIDIA doesn't have to die. None of this requires a catastrophic collapse to be a problem.

All it requires is repricing. OpenAI and Anthropic are reportedly aiming for IPO valuations north of eight hundred billion dollars apiece. Those numbers don't make sense if model capability becomes a commodity in the next twenty-four months. They make sense as monopoly rents on a defensible product. Strip the monopoly assumption, and the same companies become real businesses with real revenue, but not trillion-dollar ones.

When the labs reprice, NVIDIA reprices. When NVIDIA reprices, an index that's thirty-five percent concentrated in companies built around the AI thesis reprices with it. You don't need a crash, just a rerating. A rerating of a third of the index is a meaningful market event.

The China angle

If you were China and you wanted to attack the strongest part of the American economic stack without firing a shot, what would you do? You would not try to build a closed-source competitor, because the U.S. would just sanction you. You would do exactly what's happening. Release frontier-quality open-weight models. Publish the architectures. Subsidize the inference. Make it free, easy, and globally available to compete down the premium that American labs charge.

This is asymmetric. China doesn't need its labs to be more profitable than American labs. It needs them to be unprofitable enough, fast enough, that the American premium evaporates. The strategic objective isn't winning AI. It's de-rating the American AI trade.

Whether that's the explicit policy or just the emergent outcome doesn't matter much. The effect is the same.

What to actually do about it

For builders: architect for portability. Stop hardcoding to a single provider's API. Run a cheap open-weight model in parallel with your premium provider on real workloads and watch the quality gap close month by month. The teams that come out of this period ahead are the ones who treated their model layer as swappable from day one.

For buyers: renegotiate. The pricing power is shifting. Vendors are quietly moving from fixed pricing to usage-based pricing because they need to capture more of the value while they still can. That's a signal.

For investors: look at your index exposure honestly. If you own a market-cap-weighted S&P fund, you own the AI concentration trade whether you wanted to or not. Equal-weight indexes, international equities, and the unloved 493 are looking more interesting than they have in years. Not because the Mag 7 is going to collapse, but because the assumption that justified their multiple is.

The thing about commoditization is that it doesn't announce itself. It happens, and then everyone agrees retroactively that it was obvious. We are somewhere in the middle of that process right now.