The Myth of the Efficient-Market Hypothesis

What happened to NVIDIA and quantum computing stocks recently proves the efficient-market hypothesis ought to be rejected.

The efficient-market hypothesis (EMH) states that asset prices immediately reflect all relevant available information. Yet market inefficiencies exist due to friction in the market (e.g. transaction costs or transaction delays), information asymmetry, external economic events, or irrational behaviour. The existence of such market inefficiencies is sufficient evidence against the EMH. We must therefore reject it—if not on principle, then at least based on the recent collapses in quantum and NVIDIA stocks.

The case of quantum computing

Quantum computing is an emerging technology. Quantum advantage is at least five years away, which may even be achievable with next-generation noisy quantum computers, or NISQ machines. Fully fault-tolerant quantum computers are probably more than ten years away. Most vendors even say so. They may use marketing phrases such as “quantum realized” and “quantum utility,” but that does not change the fact that we are 5–10 years away from the quantum revolution.

When Jensen Huang, CEO of NVIDIA, a company that does not manufacture quantum computer chips at all, said it would probably take two decades rather than at least one, the stock prices of various quantum computing companies tanked. His opinion on quantum computers is as irrelevant as Mark Zuckerberg’s, who has mostly piped a few billion dollars into /dev/null (a.k.a. the metaverse) before pivoting to AI in an attempt to placate investors.

Google, D-Wave, Rigetti, SpinQ, and even a few industry experts refuted these claims in the days and weeks after, though the damage had already been done. Note that Jensen Huang and Mark Zuckerberg shared zero information with the market, merely their own opinions. If the market were truly efficient, it would have ignored them. Their opinions on quantum computing are irrelevant. The market should have also seen the statement from Huang as self-serving: NVIDIA’s GPUs, not QPUs, will be needed for at least twenty more years. Instead, the market acted as if it was privy to an epiphany. It only goes to show how misinformed investors and so-called industry experts often are.

The day after Mr Huang’s statement, NVIDIA announced a quantum day to distract from the havoc wreaked. Very classy!

The case of DeepSeek and NVIDIA

NVIDIA got a taste of its own medicine when DeepSeek released its latest LLMs into the wild. These allegedly did not require as many GPUs (from NVIDIA) as OpenAI’s or Anthropic’s models. DeepSeek could therefore train their models much more cheaply, by several orders of magnitude, in fact! And it only took DeepSeek two months to build their model. That announcement wiped almost $600b off NVIDIA’s market value.

DeepSeek described various innovations, such as distillation that enabled the company to train its AI models more efficiently, though the claims of massive cost savings were questioned by some. Instead of a little under $6m, it probably cost DeepSeek $1.6b in capital expenditure alone. Of course, DeepSeek could not publicly admit that it owned a fleet of fancy NVIDIA GPUs, because that would be tantamount to saying they had violated the chip bans. And NVIDIA would not want to admit they profited from bypassing said restrictions either.

In the case of DeepSeek and NVIDIA, the market ought to have been cautious and not take what any company says at face value. Even if we were to ignore infrastructure costs and compare the $6m from DeepSeek to the $100m in pure training costs from OpenAI for GPT-4, it is still a difference of two orders of magnitude. Such extraordinary claims require extraordinary evidence. Not reckless speculation.

The myth of the EMH

The recent drops in NVIDIA and quantum computing stocks show that the lack of technical know-how as well as the tendency to act before thinking are key reasons for so-called market inefficiencies, too. In both cases, the market reacted to either irrelevant information or unverified claims. That is not efficiency. It is stupidity.

The same can be said about the so-called breakthrough from Microsoft with the Majorana 1 chip. Both the authors of the Nature paper and the editors clearly state that the research does not prove the existence of Majorana zero modes (MZMs), which are a necessary ingredient to make topological qubits with superconducting nanowires. All the researchers demonstrated was a single-shot fermion parity measurement. That is pertinent because quantum information is stored in fermion parity, but only if MZMs exist.

Even if Microsoft’s claims are true, and they manufactured MZMs on a quantum chip, they have yet to share that evidence, and even then they are still many years away from a fully fault-tolerant quantum computer: scaling from an eight-qubit lab prototype to a million-qubit production system is far from straightforward. Once again, extraordinary claims require extraordinary evidence, not merely colourful marketing materials and vague assurances. An efficient market ought to know better, especially since Microsoft has made false claims about topological qubits before.

The only way to hold on to the efficient-market hypothesis is to see it as aspirational and asymptotic: stock prices will reflect publicly available information eventually. Maybe.