Navigating Moats and Gaps

Most of the market focus this week was on Nvidia, which released its latest quarterly results and forward-looking guidance. There’s no question that Nvidia has become the avatar of the AI boom, thanks to its status as the market-dominating supplier of GPUs and AI accelerators.

We have blown well past the point at which Nvidia became a bellwether for AI. As Nvidia goes, so goes much of the rest of the AI trade on the public markets. For multiple reasons, it won’t stay that way forever, but that’s how it is today.

We have what we can justifiably describe as an unbalanced market. When something is unbalanced, it is at perpetual risk of collapsing and falling over. When it falls, becoming prostrate or supine, we wonder whether it will be able to get back up. In the case of AI, we can reasonably assume that it possesses enough athleticism and resilience to right itself. As a market, it’s youthful and vigorous, not geriatric. Still, falls can hurt and injuries can linger.

As I mentioned earlier this week, anxiety pervades the public markets. Investors have begun to feel that an AI bubble has formed and is unsustainable. Muscle memory and momentum are keeping the AI trade climbing, but the ascent is now tentative and fitful. There’s suddenly a lack of conviction, a wariness in the markets. Just a short time ago, whenever the AI cheerleaders among the Wall Street analysts shook their pompoms, institutional investors in the stands followed suit by lustily doing the wave.

What’s behind the market’s emergent hesitancy and uncertainty?

It’s all about the money. LLMs are getting bigger and the costs of supporting them, with datacenters and all their attendant infrastructure — plus the costs of watering and fueling that infrastructure — continue to skyrocket.

There’s a surfeit of supply-side spending, escalating by the minute, and it can only be justified if demand-side revenue attributable to AI begins to materialize.

This, by the way, isn’t a question of AI’s utility. I think most of us now agree that AI will be with us for a long time. AI has a future, but what’s in question are the dimensions and scope of that future. Given the prodigious supply-side spending, AI will have to generate remarkable revenue and profit. We might reasonably wonder how spending on AI services can expand so quickly and so aggressively. For that to happen, enterprise customers will need to be confident, if not certain, of compelling ROI. Many are not there yet.

Do You Feel Lucky?

Whether you believe AI can square this challenging circle depends on your level of tech optimism and your time horizon. Still, as the AI datacenters get bigger, and as the cost of infrastructure grows, the revenue- and profit-generating bar that AI will have to clear gets higher.

A disproportionate share of investor and market focus, as I mentioned above, has been on Nvidia. Nonetheless, the other protagonists in the AI passion play are the hyperscalers, primarily Amazon, Google, Microsoft, and Meta. The first three are benefitting from the market successes of their cloud platforms, which have attracted and retained scores of enterprise workloads, including AI. The last vendor mentioned, Meta, attempts to use AI as a means of solidifying and extending its multifaceted consumer platform.

Despite all their past glories, however, hyperscalers face an AI dilemma. Let’s first consider an advantage that one might reasonably misinterpret as a disadvantage.

The high cost of supplying AI services — the massive amounts of data required for models, the costly AI expertise, the gigantic and proliferating datacenters, the exorbitantly priced GPUs and AI accelerators, the network upgrades, and the energy demands — constitute a capex moat. Very few companies can play the AI scale game. To do so, you have to have oodles of money. The cloud giants and Meta are not short of cash, and that gives them a competitive advantage in outspending everybody else in the AI universe. (I suppose Elon Musk is not short of a dollar, either, but Grok seems to misfire repeatedly, despite Dr. Muskenstein’s perpetual tinkering.)

You could say, well, that’s great for the hyperscalers. They seem to have an impenetrable moat, probably a titanium drawbridge, too. I don’t disagree, but there is a problem looming.

LLMs are resource hogs, and that’s a blessing and a curse for hyperscalers and Big Tech. It’s a blessing because only they can afford them broadly and at scale, and they can foot the bill using cashflow. Few other companies have the financial reserves to take a similar tack. But LLMs are also a curse, for the same reason that they’re a blessing: They cost a lot to run, and we can only guess at when the spending will moderate.

Investor Concerns

Institutional investors and Wall Street analysts have taken note, and they have concerns. Those concerns have grown as Meta has resorted to tapping rickety private credit and the Google, Meta, Amazon, and Oracle have issued bonds. It’s one thing to use your cash flow to fund AI buildouts, but it’s another to take on debt.

Yes, I hear what you’re saying The tech giants don’t have much debt on their books, and they can easily absorb what little they assume. Still, investors are concerned about whether debt might grow into a bigger problem as cash reserves are consumed, bond issuance proliferates, and the potential timeline for profitable AI services moves farther into the horizon.

Fortunately, when it comes to AI, LLMs aren’t the only game in town. Other approaches, include world models, are touted by some to be more likely than LLMs to achieve the holy grail of artificial general intelligence (AGI). I can’t tell you which method is more likely to attain AGI — I leave that learned debate to experts, particularly those who don’t have a horse in the race — but even if it’s a wash, if other approaches are more resource-efficient, LLMs might find themselves at a cost and pricing disadvantage. If that were to happen, the current moat engendered by vast LLMs, and the concomitant requirement for massive capex, would be turned on its head and become a liability.

Yann LeCun recently left Meta, and he’s now launching a company of his own, predicated on world models and advanced machine intelligence (AMI). Some world models (such as one on which Fei-Fei Li is working at World Labs) incorporate spatial intelligence, the ability to perceive, understand, reason about, and interact with physical spaces and objects in three dimensions. That’s what we humans do regularly, and it’s easy to see how those capabilities might enhance AI’s capacity to replicate human cognition.

My (very) limited current understanding of world models leads me to believe that they will require more, and more varied, data than LLMs. Much of that data, as with LLMs, is readily available, though LLMs often set a low bar in relation to respecting the copyright of intellectual property.

World Models: There's Work to Do

World models have technical challenges to overcome before they step into the right alongside or against LLMs. Even ardent proponents of world models (see Fei-Fei Le) concede that much needs to be done before they can be applied gainfully in scenarios such as filmmaking, game design, architecture, and next-generation robotics. Even more work is required before world models are applicable to advanced applications in healthcare, eduction, and scientific research.

The datacenters and infrastructure required for world models might be the same as, or similar to, what the cloud giants are building and deploying today for LLMs. One might assume a considerable overlap and commonality, but perhaps there will be differences. It’s conceivable that world models, if they come to pass and fulfill their promise, will require expenditures even greater than what we’ve seen for AI derived from LLMs.

Alternatively, we might see, for various specialized applications and vertical markets, models that are slimmer and more attuned to customized requirements. These could conceivably be delivered at lower costs and commensurately lower prices. Such services might reside in the hyperscalers’ datacenters, allowing the behemoths to get a piece of the action, or they might be hosted in other facilities. We’ve already seen the appearance of DeepSeek, and other open-source models from China have followed. These might be, in Microsoft’s classic parlance, “good enough” to serve a range of requirements, and the costs associated with their development and production would presumably be lower than those attached to LLMs at scale.

Martin Casado, of venture-capital firm Andreessen Horowitz (a16), said this summer that most of the AI entrepreneurs he sees in the US are using AI models devised in China. He further estimated that, of any given AI entrepreneur he sees, there is an “80 per cent chance [they are] using a Chinese open-source model.”

Chinese models, of necessity perhaps more than choice, are compelled to do more with less. Keep in mind that China has been precluded from buying and using Nvidia’s crème de la crème of AI chips. Consequently, Chinese organizations and companies had to focus more on model efficiencies while Chinese chip companies continually attempt to close the gap on Nvidia.

The upshot is that lean and mean Chinese models have made headway, and not just in China. As Casado’s comment suggests, Chinese open-source models are proving popular with AI startups, and even with larger companies, such as Airbnb, which “relies heavily’ on Alibaba’s Qwen because the Chinese model is ‘very good’ and ‘also fast and cheap.”

It’s a Complicated Game

Let me summarize the conundrum the Big Tech cloud giants confront as they invest heavily in datacenters and infrastructure to run LLMs.

First, the cost of building datacenters and procuring infrastructure is unprecedented in its enormity. Investors have taken notice, and are growing concerned both about how much hyperscalers are spending in pursuit of LLMs, especially on occasions when new debt is incurred. Even when debt isn’t incurred, investors are concerned that some of the bounty allocated to datacenters and GPUs/accelerators might be better spent elsewhere — or perhaps delivered directly to them as dividends or indirectly through stock buybacks, which can boost share prices.

As we have discussed, LLMs might not be the last word in AI, particularly if the ultimate goal is AGI. To get to AGI, the hyperscalers might have to invest even more of their cash hoard in world models, or whatever else emerges in this fast-evolving and well-funded space.

How, and when, do the cloud giants transition or load balance between LLMs and developments such as world models? Do they make the move now, at potentially greater upfront cost, or do they plan to do it later, taking the calculated risk that they can catch up and close ground if somebody else strikes first?

Then there are the more efficient, lower-cost models from China, which threaten to reduce the market shares of LLMs from OpenAI and Anthropic. The hyperscalers already run some of these models today, and they might adopt them more extensively in future. Would that allow them to reduce the footprint of datacenters and lower the cost of running the models, through, for instance, a diminution of spending on Nvidia’s premium-priced AI chips? Such a move would have the virtue of limiting the hyperscalers’ financial exposure to any disappointment in near-term enterprise demand for LLM-based AI services.

What we see on the AI landscape today might look much different two years from now. Between investor pressure and competitive threats from alternatives to LLMs, the hyperscalers confront difficult choices. What they choose to do, and at what pace, will have implications for others, including the ecosystem and supply chain that nourishes their current feeding frenzy. Competitors, too, will be strengthened or weakened by the decisions and investments the hyperscalers make.

When we look at our present reality, we often assume that the future will look much the same, but bigger and better. Perhaps that’s where things are going, but it’s equally likely that the picture will look considerably different.

A well-worn Latin proverb tells us that fortune favors the brave. That is true in many cases, but we are also well advised, as subway and train operators counsel, to mind the gap.