1. AI Circle Places “Futures Orders”, TSMC Becomes the Largest “Order Taker”
OpenAI recently made a significant move in the tech world: partnering with AMD on one hand and Broadcom on the other, it issued two “computing power futures” orders.
According to the agreement, AMD will help it build 6 gigawatts of GPUs, while Broadcom is responsible for constructing 10 gigawatts of AI accelerators and Ethernet systems, with the first batch of equipment expected to be delivered in the second half of 2026. This is no small amount — 6 gigawatts of computing power is equivalent to 600,000 high-end GPUs operating simultaneously, capable of supporting 10 large-scale AI data centers; the 10 gigawatts of accelerators could make ChatGPT’s response speed three times faster.
But who will turn this “paper computing power” into reality? The answer is clear: TSMC. As the core foundry for AMD and Broadcom, the nearly 20 gigawatts of equipment behind this order translates into a demand for tens of millions of AI chips. However, TSMC’s chairman, C.C. Wei, poured cold water on the situation: “The production capacity for AI chips is extremely tight both upstream and downstream, and it will be difficult to meet demand before 2026.”
On one side, there are urgent orders demanding delivery, and on the other, there is a production capacity that cannot be squeezed out; the leading chip foundry is indeed facing tough times.
2. Breaking Down the Orders: OpenAI Plays “Equity for Computing Power”, TSMC Bears the Hard Pressure
Outsiders see the excitement, while insiders see the intricacies. These two orders seem like a “strong alliance”, but they hide three layers of secrets, each adding pressure on TSMC:
1. Orders are “Futures”, Capacity Needs to be “Spot”
The key contradiction lies in: the orders are “futures checks”, but the capacity needs to be “immediately available”.
The collaboration between OpenAI and AMD can be described as a “capital play”: OpenAI promises to consume 6 billion watts of computing power over the next few years (the current 6 gigawatts is part of this), and AMD directly offers a warrant for 10% equity at an exercise price as low as $0.01. In simple terms, OpenAI becomes a shareholder at almost no cost and can offset procurement costs with rising stock prices; AMD, in turn, secures future orders through equity.
However, TSMC has no such “buffer cushion”. Whether it’s AMD’s Instinct MI450 GPU or Broadcom’s AI accelerator, they must be produced using TSMC’s advanced 4nm process. A 4nm production line takes two years from construction to mass production, but the orders require delivery by the end of 2026, which is akin to “asking a house that has just laid its foundation to be occupied within six months”, putting immense pressure on production capacity.
2. Demand is Too Concentrated, Capacity is “Bursting at the Seams”
AI chips have long been in high demand, and TSMC’s production lines are already insufficient.
Currently, among TSMC’s 4nm/3nm capacity, NVIDIA occupies 40%, Apple 30%, leaving little for AMD and Broadcom. The sudden addition of tens of millions of chips’ demand is like “adding dozens of tables in a fully booked restaurant”. Industry insiders estimate that just these two orders from OpenAI will require TSMC to add at least three new 4nm production lines, but each line requires an investment of over $10 billion, and they must wait for equipment to arrive and for technology debugging, which is simply too late.
Even more challenging is the “tightness in both upstream and downstream capacities” — not only is the front-end process for manufacturing chips insufficient, but the back-end capacity for packaging and testing is also in crisis. An AI chip must go through dozens of packaging processes, and TSMC’s advanced packaging capacity utilization has already exceeded 95%, leaving no room for additional orders.
3. The Industry is “Affected”, No One is Having a Good Time
TSMC’s capacity crunch has already begun to “drag down” the entire tech industry.
Smartphone manufacturers are the first to feel the chill: the delivery time for new processors originally planned to be produced using TSMC’s 4nm process has extended from three months to six months; automotive chip manufacturers are even more helpless, with some automotive-grade chips having to turn to Samsung for foundry, resulting in a 20% cost increase. Analysts joke: “Now, trying to book TSMC’s capacity is more competitive than grabbing concert tickets; not only do you have to queue, but you also have to watch their ‘facial expressions’.”
This is like “traffic congestion at a hub”; TSMC, the “heart” of the industry, is providing insufficient blood supply, causing the entire tech supply chain from smartphones to cars to smart homes to “slow down”.
3. TSMC’s “Dilemma”: To Expand or Not to Expand
Faced with this wave of “computing power frenzy”, TSMC is actually caught in an awkward “multiple-choice question”:
1. Expand? Afraid of Becoming a “Dumping Ground”
Expanding now seems to solve the immediate problem, but the risks are significant. The “bubble theory” surrounding the AI industry has never ceased; Jensen Huang has pointed out: “Some collaborations involve orders being exchanged for equity before the products are even built, which essentially shifts the risk.”
If AI demand cools after 2026, TSMC’s new production lines will become “idle assets”. It’s worth noting that during the Bitcoin craze in 2018, many chip manufacturers expanded wildly, only to see demand plummet, resulting in production line utilization rates dropping below 30%, leading to significant losses. TSMC does not want to repeat that mistake.
2. Not Expand? Lose “Voice” in the Industry
However, if they do not expand, they will have to watch customers slip away. If AMD and Broadcom cannot secure enough capacity, they may divert some orders to Samsung. Samsung has long been eyeing this opportunity and recently announced a 30% increase in its 4nm capacity, just waiting to snatch TSMC’s customers.
More critically, missing out on this AI boom could cause TSMC to lose its industry dominance. AI chips are the core battleground for the next decade; whoever controls the capacity will have the voice. If Samsung seizes the initiative, TSMC’s position as the leading foundry may be at risk.
3. Robbing Peter to Pay Paul? Offending Old Customers
Currently, TSMC’s approach is to “prioritize major customers”, but this has alienated smaller clients. Some startup chip companies complain: “We used to be able to secure small batch capacity, but now we are directly told ‘let’s talk next year’, and the company is on the verge of collapse.”
Even major clients like Apple and NVIDIA face the risk of “being diverted”. Reports indicate that the delivery time for NVIDIA’s H200 GPU has already been delayed by two months, and Apple’s A19 chip may also face production delays — if these “big players” are dissatisfied, TSMC’s situation will only worsen.
4. Cold Reflections Behind the Frenzy: AI Computing Power Must Ultimately “Land”
The collaboration between OpenAI, AMD, and Broadcom appears to be a “great leap forward in computing power”, but the underlying concerns deserve attention:
First, there are doubts about the “authenticity of the orders”. OpenAI’s revenue this year is only $4.5 billion, yet it has signed trillion-level chip orders, relying entirely on investments and equity financing from chip manufacturers to survive. This is akin to “having a monthly salary of 5,000 yet daring to order a million-dollar luxury car”; whether they can truly pay is still uncertain. If the funding chain breaks in the future, these “futures orders” could turn into “waste paper”, and if TSMC expands prematurely, it would become a “sucker”.
Secondly, there is a risk of “capacity mismatch”. AI chips have extremely high process requirements; while 4nm/3nm capacity is tight, there is surplus capacity for mature processes like 28nm. If TSMC blindly expands advanced processes, and demand shifts towards mature processes, the capacity will be severely wasted.
More importantly, the industry needs to “cool down rationally”. The current AI circle resembles a “nationwide speculation on computing power”; many companies are following the trend to order chips without considering how to use them. This is similar to the “mining craze” of the past, which ultimately left a pile of idle mining machines; such “resource waste” will eventually backfire on the entire industry.
5. Conclusion: The Feast of Computing Power Must Ultimately Be Paid for with “Strength”
OpenAI’s “big orders” have increased pressure on TSMC, fundamentally reflecting the contradiction between the “frenzy of the AI industry” and the “slow pace of chip production capacity”. On one side is the insatiable thirst for computing power, and on the other is the objective limitations of manufacturing processes; this “supply-demand game” has no shortcuts.
TSMC’s predicament also serves as a wake-up call for the industry: the development of AI cannot rely solely on “orders piled up into a bubble”; it must also depend on solid technology and rational demand. As Jensen Huang said, true collaboration should be about “selling products”, not “exchanging orders for equity”.
In the future, those who will ultimately succeed will not be the companies that “play capital games”, but the “pragmatists” like TSMC who can withstand production capacity pressures and focus on technology, and those who can turn “paper computing power” into real innovative results. After all, no matter how lively the feast of computing power is, it must ultimately be paid for with “strength” — without a solid manufacturing foundation, even the most beautiful AI dreams are just castles in the air.