Will DeepSeek deep-six the US economy?
Tech crash could turn into a funding crisis if expectations of AI revenue prospects sour due to lower-cost Chinese start-up
America has financed a current account deficit that bloated to US$1.2 trillion in 2024 by selling tech stocks to foreigners. Tech stocks, meanwhile, are trading at valuations not seen since 2000, when the NASDAQ Composite began a descent that wiped out 75% of its market capitalization by 2002.
Could a tech crash turn into a funding crisis for the United States if expectations sour on the revenue prospects of artificial intelligence? The January 27 crash in AI-related stocks in response to cheaper and better Chinese competition raises troubling questions. These questions have the undivided attention of every equity investor in the world.
Foreigners stopped buying US debt of all kinds – Treasury, mortgage, and corporate – after the post-Covid inflation of 2021 and the Federal Reserve’s consequent rise in interest rates. That denoted the end of a 40-year bull market in US bonds. From a 1981 peak of 15%, the US 30-year bond yield fell in a nearly straight line to an August 2020 low of just 1.41%.
The inflationary surge of 2021-2022 put an end to this bull run. In March 2022, moreover, the US and its allies seized half of Russia’s $600 billion in foreign exchange reserves, prompting other central banks to shift away from US Treasury securities to gold and other assets.
But the world’s appetite for American tech stocks has been bottomless for the past ten years, whetted during the past year by the advent of Large Language Models (LLMs). Are elevated valuations for AI-related stocks justified? That depends on two factors: Which sectors are likely to generate revenues from AI and how fast they will generate them.
China’s DeepSeek R1 model appears to have achieved a breakthrough in model efficiency: novel architecture and related optimizations reduce the required computation by one or two orders of magnitude.
DeepSeek, moreover, offers its model at a small fraction of the cost that its US competitors now charge. That isn’t necessarily bad for the US tech industry as a whole. If China has a better technology, US companies can adopt it rapidly, and lower costs for AI modeling will benefit the users of AI models.
There are seven major categories of AI applications in which the US and China compete. China is ahead in most of them and its AI prowess is likely to increase its lead. They are
- Manufacturing: China has poured enormous resources into factory automation. One gauge is the number of factories outfitted with dedicated 5G networks, which support AI applications. China claims 10,000 such installations, while the US has only a few dozen, concentrated in the auto industry. The advantage is strongly in China’s favor, and advances in AI are likely to enhance it. But US manufacturing has had small impact on equity valuations.
- Internet of Things: China is ahead in automating transportation and warehousing, with fully robotic warehouses now in operation.
- Robotics: China installs more industrial robots each year than the rest of the world combined and is now a major producer of industrial robots.
- China is the leader in the so-called low altitude economy, cited by government planners for the first time in a December 2024 working paper. Drone taxis, drone deliveries and other applications are now a $100 billion business in China and are expected to double by 2026.
- Autonomous vehicles: We’ll call this a toss-up between the US and China, although China already has autonomous taxi services operating on a small scale.
- Large Language Models: again, a toss-up. The gains to be harvested by LLMs include the Philippines’ $40 billion a year call center business, in which human operators can be replaced by AI systems to a considerable extent. But the possibilities for LLM applications are so varied and extensive that predictions are guesswork at this stage.
- Biotech: The US has a distinct advantage with a strong pharmaceutical development infrastructure. China has a lead in medical data, but America’s complex of large pharmaceutical companies, startups and venture capitalists give it an edge.
The big question mark over LLM monetization is timing. The payoff could be big but will probably take longer than expected to materialize.
Enterprise deployment of LLMs still has a minimal impact on corporate earnings and human adaptation (management buy-in, workflow modifications, etc.) seems to be years away. Cost savings for certain categories of expenses, such as call centers or routine coding tasks, might be realized quickly But the adaptation of AI for higher-skill work is still at a very early stage.
What does this mean for chipmakers like Nvidia? One could argue a bullish Nvidia case based on all of the AI sectors listed above, on the assumption that Nvidia GPUs will power a great deal of this activity. However, this hypothesis requires closer scrutiny of Nvidia’s competitive advantages.
Nvidia’s advantage is strongest in computation for training language and vision models, but less so for inference (running the resulting models to produce useful results). Notably, the new DeepSeek models already run quite well on Huawei’s Ascend AI chips, with similar or even better cost performance than on Nvidia H800s (the weakened Nvidia chip cleared for export to China).
If we add up all these considerations, the case for the top US tech companies (the so-called Magnificent Seven) to dominate equity returns going forward is a lot weaker than the market presently perceives it. If we are right, and tech market valuations shrink to some significant extent, what are the macroeconomic implications? More than in any period of US history, key capital flows depend upon a very small number of very big companies.
Let us suppose that foreigners were to reduce their purchases of tech stocks as valuations dissipate. To finance its current account deficit and federal budget deficit, the United States would have to sell more bonds to domestic and foreign buyers. The chart below shows the amount of new Treasury debt bought by US banks, US households, foreign official institutions, and foreign private investors, respectively.
Banks stepped in and absorbed the flood of Treasury debt that financed the $4 trillion in Covid subsidies during 2020-2021, but exhausted their store of savings deposits by 2023. The biggest increase in new investment in Treasury securities came from households, attracted by the higher interest rate on Treasuries. Foreign private buyers also increased their holdings of Treasuries by a smaller amount.
A full-blown financial crisis is most unlikely. The cash-burning dotcoms of 2000 have been replaced by cash-rich monopolies like Microsoft, Google, Apple, Amazon and Meta. The United States can adjust to an air-pocket in foreign demand for its tech stocks, for example by offering higher bond yields to domestic and international investors.
But the DeepSeek shock exposes flaws in the core strategies of Big Tech as well as the stratospheric valuation of its best-performing stocks. The outcome is likely to be persistently higher interest rates, lower growth, a reduction in wealth and stiff economic headwinds.
The S&P’s technology sector, correspondingly, trades at a P/E of 37, compared to an overall P/E for the S&P 500 of 26. That accounts for most of the gap between the lofty valuation of American stocks versus equities in Europe, Japan and China.
A brass-tacks gauge of equity valuation is the free cash flow (FCF) yield, namely the ratio of cash income to market price. The lower the FCF, the greater the expectations about future growth (investors accept less current income because they expect more income in the future). For the S&P 500 as a whole, FCF is below 3, a level not seen since the eve of the tech stock crash of 2000.
For a monopoly like Microsoft, the free cash flow yield has fallen to just 2, the lowest on record.
Big Tech doubled in capital expenditure between 2020 and 2024, and continues to make enormous commitments to data centers supporting AI. The DeepSeek shock calls into question the economic viability of these plans: If Chinese developers can build state-of-the-art models with less powerful chips, exploiting model architecture innovations, the raw computing power now under construction may be vastly overvalued.
To entice price-sensitive buyers into the Treasury market, the US government—still running a record peacetime non-recession deficit of 6% to 7% of GDP—probably will have to offer higher yields. That’s a problem for the economy and also a problem for the Treasury, which is already paying $1 trillion a year in interest, nearly quadruple the service cost of America’s national debt in 2020.
It also puts a headwind in front of the US economy for interest-sensitive activity, particularly housing. Longer-term, the US risks an Italian sort of spiral, where the rising cost of debt service eats away at the budget and limits what the federal government can do to support the economy.
Steve Hsu is professor of theoretical physics and of computational mathematics, science, and engineering at Michigan State University, and the founder of several AI startups. Follow him on X at @hsu_steve. David P Goldman is deputy editor of Asia Times. Follow him on X at @davidpgoldman