The current atmosphere in Silicon Valley feels less like a technological evolution and more like a high-stakes game of musical chairs. For the past year, the narrative has been singular: more compute equals more intelligence, and more intelligence equals infinite valuation. Developers are racing to integrate the latest frontier models, and venture capitalists are pouring billions into companies that have yet to prove a sustainable path to profitability. It is a world where a valuation jump of several hundred billion dollars can happen between two product updates, and where the mere possession of H100 clusters is treated as a sovereign reserve of wealth. But as the numbers climb toward the trillion-dollar mark, a familiar voice of skepticism is emerging to question whether the foundation is made of silicon or sand.
The Metrics of the Frontier Race
The scale of the current AI investment cycle is best illustrated by the recent capital movements surrounding Anthropic. The company recently secured a Series H funding round totaling $65 billion, pushing its valuation to approximately $965 billion. This figure is not just a milestone; it is a statement of dominance, as it surpasses the estimated $852 billion valuation of OpenAI. This surge in valuation coincides with the release of Claude Opus 4.8 on May 28, an update that arrived just 41 to 43 days after the previous Opus 4.7 version. The technical gains accompanying this release are quantifiable. In the GDP vala metric, which measures agentic capabilities, Claude Opus 4.8 recorded 1,890 ELO, placing it 121 points ahead of GPT 5.5. Furthermore, the model demonstrated significant efficiency gains, reducing output tokens by 35% and the number of steps required for complex tasks by 15%.
These efficiency gains translate directly into benchmark dominance. On the SWEBench Pro software engineering benchmark, Claude Opus 4.8 achieved a score of 69.2%, significantly outperforming GPT 5.5 at 58.6% and Gemini 3.1 Pro at 54.2%. The leap is even more pronounced in the Graph Walks 1 million token version, where the model hit 68.1%, a massive jump from the 40.3% recorded by version 4.7. While these numbers suggest a trajectory of exponential growth, the market is also shifting toward a workhorse class of models—cheaper, faster tools designed for high-volume utility. Examples include Google's Gemini 3.5 Flash and Composer 2.5. The AI code editor Cursor has leaned into this trend, releasing its own optimized coding model, Composer 2.5, which focuses on reliability in complex instruction following and long-term task persistence. Even the external testing grounds of AI Arena have seen the emergence of Gemini 3.2 Flash, which shows marked improvements in SVG generation over the standard Gemini 3 Flash available in AI Studio.
This shift toward infrastructure is also visible in the public markets. Dell has seen its stock price climb 240% this year, with an 80% surge following its last two earnings reports. Dell's strategy is not to build the models, but to build the environment they live in—providing the racks, cooling systems, and integrated NVIDIA GPU servers that make the models possible. This represents a migration of value from the chip designers to the systems integrators. Meanwhile, the ambitions of other AI-adjacent giants remain staggering. SpaceX, according to an S-1 filing submitted on May 20, reported $18.7 billion in revenue alongside a $4.9 billion net loss, yet the company is targeting a public market valuation of approximately $2 trillion.
The Brute Force Fallacy and the Infrastructure Wall
Despite the impressive ELO scores and the flurry of funding, Michael Burry sees a pattern that mirrors the financial collapses of the past. The core of Burry's skepticism lies in the method of achievement. He argues that the current progress of companies like Anthropic is based on a brute force approach—throwing astronomical amounts of computing power and data at a problem to squeeze out marginal gains. In Burry's view, this is not a sustainable moat. He posits that computing power will eventually be commoditized, much like internet bandwidth, stripping away the competitive advantage of those who simply spent the most on hardware.
This disconnect between capital expenditure and operational reality is already manifesting in the physical world. Microsoft has invested between $300 billion and $500 billion in GPUs, yet a significant portion of this hardware has reportedly sat idle in warehouses. The bottleneck is no longer the chip itself, but the surrounding infrastructure: power grids that cannot handle the load, cooling systems that cannot dissipate the heat, and wiring that cannot support the density. The belief that hardware acquisition equals competitiveness is being dismantled by the reality of electrical engineering. When GPUs cannot be plugged in or cooled, they are not assets; they are expensive paperweights.
Burry also warns against tokenmaxxing—the trend of maximizing token production as a proxy for value. He suggests that the current demand for compute is a false signal that will eventually lead to massive oversupply. The year 2025 is being viewed as a proof-of-concept phase for data centers and energy needs, but the financial concentration is precarious. With nearly 50% of the S&P 500 concentrated in the top ten stocks, investors are being pushed into private markets to find diversification, which only further inflates the valuations of companies like Anthropic and SpaceX.
There is also a psychological tension emerging within the developer community. The industry has pivoted from the practical machine learning of the 60s and 80s—which focused on statistical operations research, decision trees, and logistic regression to solve supply chain problems—to a narrative of AGI and existential risk. Burry and other critics argue that terms like AGI are primarily PR tools used to confuse the public and inflate valuations. The anthropomorphization of AI—the idea that a model understands or thinks—is viewed as a science-fiction distraction that obscures the reality that these are statistical engines. This narrative shift is reportedly demoralizing a new generation of developers, aged 20 to 25, who find the apocalyptic warnings of AI leaders to be a deterrent to genuine technical contribution.
Even the benchmarks themselves are under scrutiny. There are growing concerns that models are being tuned specifically to score higher on tests rather than becoming more capable in the real world. While Google is internally testing Remy, a 24/7 personal AI agent integrated into Gmail, Docs, and Calendar to handle complex workflows, the real test will be whether these agents can solve edge cases and eliminate hallucinations in legacy codebases—problems that the CEOs of Cursor and Cognition admit remain unsolved despite the increased efficiency of tool calling.
This cycle of disruption is not new to the financial world. High-yield bonds, leveraged loans, and ETFs were all born from clean-sheet thinking that challenged existing business fundamentals. However, the risk today is what Burry describes as the difference between a heart attack and cancer in financial services. A heart attack is a sudden liquidity crisis caused by borrowing short-term to lend long-term, similar to the collapse of Drexel in 1990. The AI bubble, by contrast, may be a slower, systemic failure of value—a cancer where the cost of maintaining the intelligence exceeds the economic value it generates.
As the market moves past the initial awe of fluent language generation, the cold logic of the balance sheet is returning. The era of valuing a company based on the size of its GPU cluster is ending, replaced by a demand for actual output efficiency and sustainable margins.
