![[ One AI, Two Worlds ]Meta Released Glimmer Today. The Market Barely Reacted—and That May Tell Us What Wall Street Is Ac...](https://img4.financial1000.com/364/309/1516168143643095.jpg)
10/08/2026
[ One AI, Two Worlds ]
Meta Released Glimmer Today. The Market Barely Reacted—and That May Tell Us What Wall Street Is Actually Willing to Pay For.
Meta made two important moves today. First, Mark Zuckerberg published a roughly 6,500-word manifesto laying out his vision for the future of AI. Second, Meta introduced Muse Glimmer, an open-weight model with roughly 30 billion parameters.
What makes Glimmer interesting is not that it is “the most powerful model in the world.” It is that it is small and efficient enough, at least theoretically, to run on a high-end personal computer or workstation. In other words, Zuckerberg’s larger message is that the future of AI should not be concentrated entirely inside the massive data centers of a handful of companies. Eventually, individuals may have their own AI running on their own computers, and that intelligence could move beyond PCs into cars, robots, drones, and all kinds of physical devices.
Yet Meta’s stock barely reacted to the announcement today. That is a striking contrast with what happened a few months ago when Meta introduced Muse Spark.
Back in April, Meta released Muse Spark and the market was genuinely excited. Meta shares jumped sharply that day. At the time, one of Wall Street’s biggest questions about Meta was fairly straightforward: after spending so much money building AI data centers, buying GPUs, and recruiting some of the world’s best AI researchers, could Meta actually produce a world-class frontier model? Muse Spark gave investors evidence that Meta was still competing at the front of the AI race. It showed that the company could still go head-to-head with OpenAI, Anthropic, and Google, and the market rewarded Meta accordingly.
Today, Zuckerberg published a sweeping AI manifesto and introduced Glimmer, yet the market reaction was remarkably quiet. Why? I think the market is subtly telling us something: the AI Wall Street is willing to pay for today is still centralized, frontier-level intelligence. Personal AI may have a tremendous future, but at least for now, that is not where the money is.
We often talk about AI as though there is only one race. Increasingly, however, I think AI is developing along two very different paths.
The first is centralized AI: extremely large models and potentially superintelligent systems running inside enormous data centers. OpenAI, Anthropic, Google, Meta, and xAI are still fighting their most intense battles in this arena. These models require huge numbers of GPUs, tremendous amounts of electricity and memory, and extraordinarily expensive infrastructure. In return, they can perform increasingly difficult tasks: coding, research, data analysis, software operation, scientific discovery, enterprise agents, and eventually sophisticated government and defense applications.
The second path is about taking AI out of the data center and putting intelligence directly into computers, cars, robots, drones, and other physical machines. Google’s Gemma family represents part of this direction, and Meta’s Glimmer clearly points toward the same future.
Both directions are important. But from an investment perspective, the question is different: which one is actually making money today? At the moment, the answer seems quite clear. It is the first one.
Enterprise AI already has a much easier return-on-investment calculation. That is one reason Anthropic has become so important. Claude is not valuable simply because it is a good chatbot. It is particularly strong in coding, enterprise work, and agentic tasks, and businesses are willing to pay for those capabilities. If a corporation spends $20 million a year on AI but can save $50 million in labor costs or substantially increase the productivity of its engineers, the investment is relatively easy to justify.
Now imagine telling investors that, someday, everyone will have a personal AI running inside their computer. Most investors would probably agree that this sounds plausible. But the next question immediately becomes: who makes the money? Meta? Microsoft? NVIDIA? AMD? Intel? Cloud or telecom operators? The robot manufacturer? Right now, the answer is not particularly clear. For Wall Street, personal AI is therefore an attractive vision of the future, but it is not yet something that can easily be plugged into an earnings model.
This brings us to another term that has become increasingly important: MoE, or Mixture of Experts. Why has MoE suddenly become such an important part of the AI discussion? The basic concept is actually fairly simple. Imagine an AI model containing one trillion parameters. In a conventional large model, a very large portion of those parameters may have to participate in computation. An MoE model works differently. It might contain one trillion total parameters but activate only something like 30 billion of them for a particular token—and importantly, it does not activate the same 30 billion every time.
Inside the model is a router. As the AI processes each token, that router decides which “experts” should handle it. A programming-related token might activate one combination of experts, a mathematical problem another, and Chinese language processing another. The entire model can therefore be enormous, while only a relatively small part of it is actively computing at any particular moment. The biggest advantage is straightforward: less computation, which potentially means fewer GPU cycles, less electricity, and lower cost per unit of intelligence.
But MoE does not necessarily mean that data centers will become smaller. The intuitive argument is that if AI becomes five times more efficient, perhaps we will need only one-fifth as many GPUs. That may turn out to be completely wrong because of an old economic phenomenon called the Jevons paradox: when something becomes cheaper and more efficient, people often end up consuming much more of it.
Suppose the cost of AI falls fivefold. A corporation may not say, “Wonderful, now we can cut our AI budget by 80%.” Instead, it may decide that AI is finally cheap enough to put agents into customer service, finance, engineering, legal work, marketing, supply-chain management, cybersecurity, and research. An engineering department might eventually give every engineer several AI agents. The cost of each unit of intelligence could fall dramatically while total consumption rises tenfold, fiftyfold, or even a hundredfold. The result could be that data centers become busier, not quieter.
Agents can magnify this effect even further. Today, when we ask ChatGPT a question, we may think of it as one task. A future agent could be very different. You might simply tell it, “Find the problem in this software.” Behind the scenes, however, that agent could read 20 files, call a model 30 times, modify code, run tests, discover an error, rewrite the code, test it again, and then ask another model to verify the result before finally returning an answer.
You see one task. The data center may see dozens or hundreds of inference operations. That is why the simple argument that “AI models are becoming more efficient, therefore NVIDIA will eventually sell fewer GPUs” may be deeply misleading. The opposite could happen: the more efficient AI becomes, the more intelligence we consume.
This also helps explain why Chinese AI companies have become so aggressive in MoE and other efficiency technologies. To some extent, geopolitics has forced the issue. DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, and MiniMax are all important participants in this efficiency race. China certainly did not invent MoE; Google was researching sparse expert architectures years ago, and Europe’s Mistral introduced Mixtral relatively early in the open-model era. But Chinese AI companies operate under a special constraint: U.S. export restrictions have limited their access to the most advanced AI chips.
That means American and Chinese AI companies sometimes approach the same problem from different starting points. An American company may be able to ask, “How do I get more GPUs?” A Chinese company may be forced to ask, “I only have this much compute. How do I squeeze more intelligence out of every GPU I have?”
That pressure encourages aggressive work on MoE, quantization, model distillation, more efficient attention mechanisms, and inference optimization. In other words, U.S. semiconductor restrictions were intended to create a weakness for Chinese AI development, and they certainly remain a significant constraint. But those same restrictions have also pushed Chinese companies to confront a problem the entire AI industry will eventually have to solve: how do we produce more intelligence using less electricity and less compute? That constraint could, paradoxically, become a source of competitive strength.
This is also why I do not think Zuckerberg’s emphasis on open AI and open weights is simply idealism. There is a strong geopolitical dimension to it. Imagine a future in which the global open-model ecosystem is increasingly built around DeepSeek, Qwen, Kimi, and MiniMax. American companies could find themselves in an uncomfortable situation where the strongest proprietary frontier models are still American, but the open models that developers around the world actually download, modify, deploy, and embed into their products increasingly come from China.
Meta clearly does not want that future. Glimmer can therefore also be interpreted as a statement that the United States cannot simply surrender the open-model ecosystem to China.
So where do dense models fit into all of this? Returning to Glimmer, I do not think the most interesting future for a roughly 30-billion-parameter model is necessarily to compete directly with a trillion-parameter data-center model. Its real opportunity may be to bring AI into the physical world.
Consider an advanced autonomous drone. A drone cannot ask a distant data center about every decision it needs to make. The network may disappear, bandwidth may be poor, and in a defense environment communications may even be intentionally disrupted. Yet the drone still has to see, understand its surroundings, avoid obstacles, determine its position, react to changes, and complete its mission. Those decisions need to happen on the machine itself. What matters in that environment is not simply having the world's largest model. The AI needs to be smart enough, small enough, power-efficient enough, inexpensive enough, and fast enough to react locally.
The future architecture could therefore look something like this: enormous data centers train the most powerful AI systems, those capabilities are then distilled and compressed into 30-billion-parameter models or something much smaller, and those specialized models eventually move into robots, automobiles, drones, factories, and other physical systems.
That could create an enormous future for dense models, but the important point is that this future may not arrive nearly as quickly as the centralized AI economy. That is one reason I find Meta’s muted stock reaction today worth watching. Glimmer may be important. Zuckerberg’s manifesto may even describe what AI looks like ten years from now. But capital markets do not primarily price ten-year philosophical visions. They want to know how the hundreds of billions of dollars being invested in AI today turn into revenue next year and the year after.
For now, the answers are still enterprise AI, agents, coding, scientific research, government, defense, and cloud services—in other words, large-scale centralized AI. That is why Meta’s next major frontier model may ultimately matter more to investors than Glimmer.
If the next generation of Muse Spark can clearly demonstrate that Meta belongs at the absolute frontier alongside OpenAI, Anthropic, and Google—or even surpasses some of them—investors can immediately connect that achievement to Meta’s enormous AI capital expenditures: data centers lead to frontier models, frontier models power agents and Meta’s own products, those capabilities create API and enterprise opportunities, and those opportunities eventually produce revenue.
The road from Glimmer to revenue is much longer and less certain: data centers lead to large models, those models are distilled into Glimmer, Glimmer moves onto personal computers and eventually robots, drones, and automobiles—and then we still have to ask who ultimately captures the profit.
That is why, looking at the AI investment landscape over the next two or three years, I still believe the largest pool of money will remain inside the data center. OpenAI, Anthropic, Google, Meta, and xAI will continue competing to build the strongest centralized frontier models. MoE will also become increasingly important because it offers a way to increase overall model capacity while activating only a fraction of those parameters, potentially reducing inference cost, GPU computation, and electricity consumption. DeepSeek, Qwen, Kimi, and MiniMax have already pushed this efficiency competition aggressively, and American AI companies cannot afford to ignore it. As enterprise agents proliferate, cheaper and more efficient models may not reduce demand for data-center computing at all; they may cause total demand to explode.
The recent changes at Google can also be viewed through this same lens. Jeff Dean’s departure and the broader leadership changes at Google DeepMind naturally raise questions about what happens next to Gemma, which remains important to Google’s open-model, local-AI, and edge-computing strategy. But from the perspective of Google’s competitive position and investment value over the next few years, I think the next Gemini—and perhaps especially the next generation of Gemini Flash—may matter far more than the next Gemma.
What Google most urgently needs to demonstrate is that it remains at the very front of the centralized frontier-AI race, particularly as Anthropic gains momentum in coding, agents, and enterprise AI. Whether Google eventually describes the next Flash architecture publicly as MoE is almost beside the point. The real question is whether Google can use less active computation to produce intelligence that approaches or exceeds the best frontier models, combining model capability with superior inference economics.
That makes Google's situation remarkably similar to Meta's. Glimmer is interesting, but Meta’s next Muse Spark may be more important to the investment story. Gemma is interesting, but Google’s next truly powerful Gemini—the model that makes the market feel that Google has pushed the frontier forward again—may be much more important. In other words, the market is not yet primarily asking who can squeeze a 30-billion-parameter model into a personal computer. It is asking who can build the strongest and most efficient centralized intelligence and deploy it economically across millions of enterprise agents.
Dense models are not going away. Gemma, Glimmer, and similar models could eventually become enormously important when robotics, autonomous drones, self-driving vehicles, and industrial AI reach mass deployment. At that point, models will need more than intelligence. They will need to be compact, inexpensive, power-efficient, responsive, and capable of operating without a constant network connection. That may ultimately become the enormous market for dense local AI.
So I do not see centralized AI and distributed AI as two futures in which one eventually destroys the other. They look more like two stages of the same AI revolution. The first stage is to build the most powerful intelligence in the world and use MoE or other technologies to make that intelligence efficient and affordable enough for massive enterprise deployment. The second stage is to compress, distill, and distribute that intelligence until it can live inside every computer, automobile, robot, drone, and industrial machine.
A few months ago, Meta released Muse Spark and the market rallied. Today, Zuckerberg published his vision for the future of AI and introduced Glimmer, and the market barely reacted. Meanwhile, Jeff Dean’s departure and the leadership changes at Google DeepMind make me less interested in what the next Gemma looks like than in whether the next Gemini can once again make the market sit up and say that Google has moved the frontier.
For now, Wall Street’s message to the AI giants seems fairly simple: First show me that you can build the most powerful AI—and that you can turn it into real money. How you eventually put that intelligence into every computer, every car, and every robot is a conversation we can have later.
![[When Being Right Isn’t Enough]A recent Wall Street Journal article, titled “His Wedding Guests Were Arriving—Just as Hi...](https://img4.financial1000.com/774/175/1508514877741755.jpg)
![[Is AI Finance Ready]Three AI Security Incidents to rethink something much bigger than cybersecurity.According to recent...](https://img4.financial1000.com/451/131/1507485984511311.jpg)
