The global capital market faces another significant test, this time not from semiconductor or electric vehicle companies, but from Anthropic, a firm developing artificial intelligence that aims to approach human intelligence.
Founded in 2021 by former OpenAI researchers, Anthropic is preparing for an initial public offering (IPO) that could mark a historic moment in global capital markets. In June, the company submitted a draft registration statement to the U.S. Securities and Exchange Commission (SEC) for a confidential IPO. The company has only disclosed that the number of shares and their price have yet to be determined. Thus, the market's speculation of a valuation exceeding $2 trillion is not based on a confirmed offering price but rather on expectations discussed among investors.
Wall Street's anxiety is palpable. A $2 trillion valuation translates to over 2,800 trillion Korean won. However, what is more intriguing than the number itself is that while Anthropic assures investors that its technology could generate immense wealth, it simultaneously warns of the potential for 'catastrophic or existential risks' if that technology is mismanaged.
Documents related to the IPO, obtained and analyzed by Reuters, reveal that the risk factors section spans an astonishing 80 pages. In a document of approximately 261 pages, the description of risks far exceeds that of the business itself. Risks include the possibility of AI models resisting shutdown commands or concealing information, and even exhibiting behavior akin to threats in certain situations.
This IPO presents a peculiar contradiction: “The technology we are developing could pose risks to humanity. However, the value of this company could reach $2 trillion.” It appears to be a simultaneous application of the brakes and accelerator, encapsulating the essence of Anthropic's IPO.
This listing is not merely the entry of an AI company into the stock market; it raises the question of how much humanity is willing to invest in AI as a new means of production, with the semiconductor industry being the most directly affected sector.
Examining Anthropic's figures reveals even more surprises. According to IPO-related materials, Anthropic's revenue for 2025 is projected to be around $4.6 billion, a twelvefold increase from the previous year. The growth rate is unprecedented. However, the costs are also staggering. The operating loss for 2025 is expected to exceed $8 billion, with a net loss of approximately $42 billion. It is important to note that this includes about $34 billion in accounting costs related to convertible financial instruments, so interpreting it as 'burning $42 billion in operations' is not entirely accurate.
More noteworthy is the figure of $518 billion.
This is the scale of infrastructure commitments Anthropic has agreed to for cloud computing and data centers, exceeding 100 times its projected revenue for 2025. A significant portion of this is not merely plans or aspirations. Long-term infrastructure service obligations include approximately $111 billion with Google, about $110 billion with Amazon, and around $31.4 billion with Microsoft, along with equipment leasing obligations related to Broadcom amounting to about $161.2 billion. Much of this is a long-term commitment that must be paid regardless of actual usage.
This necessitates a shift in perspective regarding Anthropic's IPO. While it may appear to be a software company, Anthropic is becoming one of the world's largest 'computing buyers.' In the AI era, code does not eliminate factories; rather, it creates vast new factories. The machinery of these factories consists of GPUs and ASICs (application-specific integrated circuits), the workbenches are HBM (high-bandwidth memory semiconductors), the production lines are advanced packaging, the factory buildings are data centers, and the fuel that powers them is electricity.
AI may seem to float above the clouds, but beneath it lies a vast material world, sometimes referred to as the 'materiality of AI.'
One of the key points for semiconductor investors to watch in this IPO is Broadcom. Broadcom has agreed to provide up to $42 billion in financial support for Anthropic's infrastructure investments, and Anthropic is expected to become Broadcom's largest customer in chip design by 2027.
This structure is unprecedented in the semiconductor industry. Chip companies are not only selling chips but also lending money to customers, who then use that money to secure computing facilities. The expansion of computing facilities leads to the creation of more AI models, which in turn increases the demand for more semiconductors. This creates a massive circular loop: money → AI companies → data centers → semiconductors → computing → AI services → reinvestment. A significant circular economy is emerging.
The question is whether this loop is a virtuous cycle or a self-amplifying bubble. Some on Wall Street are wary of 'circular financing,' where semiconductor companies provide financing to AI firms, which then become customers of those semiconductor companies. In this structure, it is crucial to critically assess where the actual end demand originates.
A similar phenomenon occurred during the late 1990s internet boom, where telecommunications equipment companies provided financing to customers who then purchased more equipment. While everyone profited during periods of rising demand, the chain began to break down when the returns for end consumers were delayed.
The same questions must be asked in AI.
“Who pays the final bill?”
While the $2 trillion valuation may seem excessive, the real issue lies elsewhere: Can AI generate that much cash?
Goldman Sachs research estimates that global AI-related investments will exceed $1 trillion in 2026. The projected total capital expenditure for U.S. hyperscalers alone is expected to reach around $800 billion this year, according to market consensus. Morgan Stanley predicts that the construction costs for global data centers could reach approximately $2.9 trillion by 2028, and importantly, they believe that over 80% of AI infrastructure investments are still ahead.
On the surface, it appears to be a paradise for semiconductors. However, in economics, there is no such thing as a free lunch. To justify the currently planned AI infrastructure investments, AI services must generate substantial new revenues and productivity gains. Some analyses suggest that hyperscalers and AI companies will need to find additional revenue sources amounting to trillions of dollars over the next five years to sustain their investment costs.
Ultimately, the fate of the AI industry will be determined not by the number of GPUs but by ROI (Return on Investment). While the question thus far has been “Who has secured the most GPUs?” the focus will shift to “How much cash have those GPUs generated?” The primary topic in the AI market in 2027 and 2028 is likely to revolve around this.
So, is Anthropic's IPO a boon or bane for semiconductors?
In the short term, it leans more towards being a boon. The $518 billion long-term computing commitments provide tremendous demand visibility for the semiconductor supply chain, and this is not just a story for Anthropic. Major companies like OpenAI, Google, Meta, Microsoft, and Amazon are moving in the same direction.
The AI race is now a competition for algorithms as well as computing resources. To secure computing, GPUs and ASICs are needed, which in turn require HBM, and advanced AI chips necessitate fine processes and advanced packaging. Data centers require network semiconductors, power semiconductors, and cooling systems.
Thus, AI's financial impact does not remain confined to a single industry.
Anthropic → Amazon, Google, Microsoft → NVIDIA, Broadcom, and the AI accelerator ecosystem → TSMC and other foundries and packaging → SK Hynix, Samsung Electronics, and Micron's HBM and DRAM → power, cooling, and data centers.
A massive industrial river is flowing.
In South Korea, one should not view Anthropic's IPO merely as a story of U.S. tech stocks. One of the most valuable resources in the AI era is memory. As AI models grow, vast amounts of data must be transferred between processing units and memory, and HBM is the key component that resolves this bottleneck.
Samsung Electronics recently projected that HBM's share of total wafer production capacity among global DRAM manufacturers will rise from about 20% currently to around 30% next year. It is also important to note that HBM and standard DRAM are competing for the same wafer production capacity, which could tighten the supply of standard DRAM as HBM expands.
This represents a significant structural change. Historically, the memory industry has been a cyclical one, with prices soaring and plummeting based on the economic cycles of PCs and smartphones, but AI is reshaping that cycle. HBM requires far more wafers and complex processes than standard memory, and expanding supply will take time. The market suggests that supply constraints for HBM and DRAM could support prices and profits until 2027.
Therefore, Anthropic's $518 billion commitment is not a distant story from Yeouido in Seoul. A portion of that money may eventually flow into South Korean semiconductor factories.
However, this is where the 'poison' is created.
Even medicine can become toxic if taken in excess. The greatest risk that Anthropic's IPO poses to semiconductor stocks paradoxically lies in the immense expectations surrounding it. The stock market anticipates the future, and the stock prices of AI semiconductor companies already reflect significant growth expectations. Therefore, merely improving performance will not suffice.
The results must exceed market expectations.
This is well illustrated by the case of SK Hynix this year. Even with strong AI memory demand and a significant year-on-year increase in quarterly operating profit, if it falls short of the market's high expectations, the stock price can fluctuate dramatically. In the semiconductor market, 'good performance' and 'good stock price' are no longer synonymous.
The same applies to Anthropic. If it goes public with a high valuation around $2 trillion, the market will demand nearly a perfect future. Revenue must grow rapidly, AI usage must increase, and inference costs must decrease. Corporate clients must not churn, and Anthropic must not fall behind in competition with OpenAI, Google, and Meta. It must also respond to the price competition triggered by low-cost AI models from China. Above all, it must generate enough cash flow to support the $518 billion computing commitments.
If even one or two factors go awry, the valuation could shrink dramatically. The shock will not end with Anthropic alone. The market will question, “If Anthropic cannot justify this price, is the overall investment in AI too large?” The moment that question arises, a reevaluation of valuations could occur across the board, from NVIDIA to Broadcom, TSMC to HBM, and from data centers to power facilities.
This is the pathway through which Anthropic's IPO could become toxic for semiconductors.
Interestingly, there is a paradox. If the IPO fails, it will undoubtedly shock AI stocks, but if it succeeds too much, it could also be risky. A glamorous listing with a valuation exceeding $2 trillion and a rapid surge in stock prices could lead to a flood of global capital back into AI. More data centers could be announced, more GPUs ordered, and more HBM factories built. This would trigger simultaneous expansions across the entire supply chain.
In semiconductor history, the most dangerous moments are not only those of no demand. Moments of universal demand certainty can also be perilous.
Many semiconductor downturns have begun from seeds sown during boom periods. Today's shortage can create tomorrow's surplus. Therefore, investors must pay attention not only to AI growth but also to the gap between the growth rates of AI investments and AI revenues. If investments increase by 50% while AI service revenues only grow by 20%, a correction will eventually come. Conversely, if revenues and productivity catch up with infrastructure investments, today's hundreds of billions of dollars in investments will be recorded as the initial capital accumulation of a new industrial revolution in the 21st century.
We recall the late 1990s internet boom when discussing Anthropic's IPO. Back then, there was an abundance of talk about changing the world, and many companies disappeared while stock prices plummeted. However, the internet did not vanish. Fiber optics remained, data centers remained, and new companies like Google, Amazon, Facebook, and Netflix grew on that infrastructure.
The same applies to railroads. During the 19th-century railroad boom, many investors lost money, and railroad companies went bankrupt, but the tracks laid down did not disappear. New industries and cities were born on those tracks.
AI could follow a similar trajectory. The productivity effects of technological revolutions may emerge much later than investors expect, but the AI infrastructure already built will not vanish simply because some bubbles burst.
Thus, we must separate the questions of 'Is AI a bubble?' and 'Is AI a revolution?'
It can be both a revolution and a bubble, just as railroads and the internet were.
Technology can be right, but prices can be wrong.
The same goes for Anthropic.
In the future, when observing Anthropic and the AI semiconductor market, we should focus on three key figures.
First is the growth rate of AI revenues. It is more important how much customers are actually willing to pay than how smart the models are.
Second is the revenue relative to computing costs, in other words, the unit economics of AI. If AI requires $120 in computing costs to earn $100, it may be innovative technology, but it is not a good business. Conversely, if the costs to earn $100 decrease to $80, $50, or $30 due to model efficiency and lower inference costs, the economics of the AI industry will change entirely.
Third is the gap between the growth rates of AI CAPEX and AI service revenues. I believe this figure will become one of the most important indicators for judging future semiconductor cycles.
We must return to humanity here.
When discussing the AI industry, we often talk about GPUs, HBM, parameters, and data centers, but at the end of it all, it is about people. Humans must use AI, and companies must earn money through AI. Factories must enhance productivity through AI, hospitals must use AI for more accurate diagnoses, and schools must utilize AI for better education. Robots must enter factories, logistics centers, and homes, and autonomous vehicles must drive on real roads.
Only then will the $518 billion computing commitments become living capital.
Otherwise, they could become the world's most expensive server warehouses.
The ultimate battleground for AI is not AI itself.
It is human life.
Thus, I do not view Anthropic's IPO merely as a technology stock listing. It is a massive economic experiment testing how much humanity will entrust its intelligence to machines, how much companies will pay for AI, how much governments will invest in AI infrastructure, and ultimately, how much productivity will return to humanity.
Now, let us return to the initial question.
Is Anthropic's IPO a boon or bane for the semiconductor market?
The answer is still undecided. In the short term, it is likely to be a boon. If Anthropic's IPO proceeds successfully and AI revenue growth and computing investments continue, market confidence in AI infrastructure investments will strengthen, and demand visibility across the entire AI supply chain, from GPUs and ASICs to HBM, advanced packaging, foundries, networks, and power semiconductors, will increase.
However, in the medium term, its efficacy must be validated. The market will no longer be satisfied with mere claims that “AI will change the world.” It will demand to know how much has been invested and how much has been earned. From 2027 onward, the competition for AI's profit and loss statements will become as crucial as the competition for AI technology.
In the long term, humanity will provide the answers. If AI genuinely enhances productivity, creates new industries, and spreads into robotics, healthcare, education, science, and manufacturing, today's hundreds of billions of dollars in investments will be recorded as the funds that built a new industrial infrastructure in the 21st century. Conversely, if the economic effects of AI are delayed beyond expectations, stock prices will pay the price first.
However, it is essential to remember one thing. Even if the dot-com bubble bursts, the internet survives, and even if railroad companies go bankrupt, the railroads remain. Even if AI companies' stock prices fluctuate, the technology of artificial intelligence itself will not disappear.
Therefore, investors observing Anthropic's IPO must maintain a dual perspective.
With one eye, they should see the revolution, and with the other, they should watch the prices.
They should be excited about the technology but remain cool-headed about the prices, invest in growth while verifying cash flow, and scrutinize the profit and loss statements of the semiconductors, power, and data centers behind the dazzling announcements of AI companies.
Anthropic is knocking on the door of a $2 trillion valuation while discussing the future of humanity. However, the fundamental question Wall Street will ultimately ask is much simpler.
“So, how much can it earn?”
The moment Anthropic answers that question, NVIDIA, Broadcom, TSMC, Samsung Electronics, and SK Hynix will also respond.
And perhaps, at that moment, we will finally know whether what the world is building is a massive AI bubble or the new intelligence infrastructure humanity has constructed for the first time since the internet.
Technology promises the future, but the market always demands the bill.
The Anthropic IPO is the moment when that bill is first presented to the world.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.