Feed: Engadget is a web magazine with obsessive daily coverage of everything new in gadgets and consumer electronics
Feed Link: https://www.engadget.com/
---

Title: OpenAI's recent chip deals heap more pressure on TSMC

Link: https://www.engadget.com/computing/openais-re...

In recent weeks, OpenAI has signed blockbuster deals with AMD and Broadcom to
build vast numbers of AI chips. Much of the focus has been on the financial
implications, since OpenAI will need hundreds of billions of dollars to make
good on its promises. As important as it is to look at the quite implausible
financials, we also need to look at the broader implications for the
industry. Like, the chips themselves, what that spells for the AI industry as
a whole, and the added pressure on TSMC, the only chip company that can
actually build this stuff.

The Deals

OpenAIΓÇÖs deal with AMD will see the chip giant build out 6 gigawattsΓÇÖ
(GW) worth of GPUs in the next few years. The first 1 GW deployment of
AMDΓÇÖs Instinct MI450 silicon will start in the back end of 2026, with more
to come. AMDΓÇÖs CFO Jean Hu believes that the partnership will deliver
ΓÇ£tens of billions of dollars in revenueΓÇ¥ in future, justifying the
complicated way the deal is funded.

Meanwhile, BroadcomΓÇÖs deal with OpenAI will see the pair collaborate on
building 10 gigawattsΓÇÖ worth of AI accelerators and ethernet systems that
it has designed. The latter will be crucial to speed up connections between
each individual system in OpenAIΓÇÖs planned data centers. Like the deal with
AMD, the first deployments of these systems will begin in the back half of
2026 and is set to run through 2029.

Phil Burr is head of product at Lumai, a British company looking to replace
traditional GPUs with optical processors. HeΓÇÖs got 30 years experience in
the chip world, including a stint as a senior director at ARM. Burr explained
the nitty-gritty of OpenAIΓÇÖs deals with both Broadcom and AMD, and what
both mean for the wider world.

Burr first poured water on OpenAIΓÇÖs claim that it would be ΓÇ£designingΓÇ¥
the gear produced by Broadcom. ΓÇ£Broadcom has a wide portfolio of IP blocks
and pre-designed parts of a chip,ΓÇ¥ he said, ΓÇ£it will put those together
according to the specification of the customer.ΓÇ¥ He went on to say that
Broadcom will essentially put together a series of blocks it has already
designed to suit the specification laid down by a customer, in this case
OpenAI.

Similarly, the AI accelerators Broadcom will build are geared toward more
efficient running of models OpenAI has already trained and built ΓÇö a
process called inference in AI circles. ΓÇ£It can tailor the workload and
reduce power, or increase performance,ΓÇ¥ said Burr, but these benefits would
only work in OpenAI's favor, rather than for the wider AI industry.

I asked Burr why every company in the AI space talks about gigawatts worth of
chips rather than in more simple numbers. He explained that, often, itΓÇÖs
because both parties donΓÇÖt yet know how many chips would be required to
meet those lofty goals. But you could make a reasonable guess if you knew the
power draw of a specific chip divided by the overall goal, then cut that
number in half, then remove an extra 10 percent. ΓÇ£For every watt of power
you burn in the chip, you need about a watt of power to cool it as well.ΓÇ¥

In terms of what OpenAI gets from these deals, Burr believes that the startup
will save money on chips, since thereΓÇÖs ΓÇ£less marginΓÇ¥ from making your
own versus buying gear from NVIDIA. Plus, being able to produce custom
silicon to tailor the work to their needs should see significant speed and
performance gains on rival systems. Of course, the next biggest benefit is
that OpenAI now has ΓÇ£diversity in supply,ΓÇ¥ rather than being reliant on
one provider for all its needs. ΓÇ£Nobody wants a single supplier,ΓÇ¥ said
Burr.

The Factory

Except, of course, OpenAI may be sourcing chips from a variety of its
partners, but no matter whatΓÇÖs stamped on the silicon, it all comes from
the same place. ΓÇ£IΓÇÖd be very surprised if it wasnΓÇÖt TSMC,ΓÇ¥ said Burr,
ΓÇ£IΓÇÖm pretty sure all of the AI chips out there use TSMC.ΓÇ¥ TSMC is short
for Taiwan Semiconductor Manufacturing Company which, over the last decade,
has blown past its major rivals to become the biggest (and in many cases
only) source of bleeding-edge chips for the whole technology industry. Unlike
historic rivals, which designed and manufactured their own hardware, TSMC is
a pure play foundry, only building chips designed by others.

Interior at one of TSMC's FabsTaiwan Semiconductor Manufacturing Co. Ltd.

Gil Luria is Managing Director at head of technology research at investment
firm DA Davidson. He said that TSMC isnΓÇÖt just a bottleneck for the western
technology industry, but in fact is the "greatest single point of failure for
the entire global economy.ΓÇ¥ Luria credits the company with an impressive
expansion ΓÇ£considering it has had to ramp the production of GPUs tenfold
over the last three years.ΓÇ¥ But said that, ΓÇ£in a catastrophic scenario
where TSMC is not able to produce in Taiwan, the disruption would be
significant.ΓÇ¥ And that wonΓÇÖt just affect the AI world, but ΓÇ£mobile
handset sales as well as global car sales.ΓÇ¥

TSMC supplanted Intel for a number of well-documented reasons, but the most
relevant here is its embrace of Extreme Ultraviolet Lithography (EUV). ItΓÇÖs
a technology that Intel had initially backed, but struggled to fully adopt,
allowing TSMC to pick it up and run straight to the top. EUV produces the
headline-grabbing chips used by pretty much everyone in the consumer
electronics world. Apple, Qualcomm, NVIDIA, AMD (including the SOCs inside
the PS5 and Xbox) all use TSMC chips. Even Intel has been using TSMC
foundries for some consumer CPUs as it races to bridge to gulf in
manufacturing between the two companies.

ΓÇ£TSMC is the current leader in advanced 3 nanometer (nm) process
technologies,ΓÇ¥ said University of Pennsylvania Professor Benjamin C. Lee.
The companyΓÇÖs only meaningful competitors are Intel and Samsung, neither of
which pose a threat to its dominance at present. ΓÇ£Intel has been working
for a very long time to build a foundry business,ΓÇ¥ he explained, ΓÇ£but has
yet to perfect its interface.ΓÇ¥ Samsung is in a similar situation, but
Professor Lee explained it ΓÇ£has been unable to attract enough customers to
generate a profitable manufacturing business.ΓÇ¥

Professor Lee said that TSMC, by comparison, has become so successful because
of how good its chips are, and how easy it is for clients to build chips with
its tools. ΓÇ£TSMC fabricates chips with high yield, which is to say more of
its chips emerge from the fabrication process at expected performance and
reliability.ΓÇ¥ Consequently, it should be no surprise that TSMC is a money
making machine. In the second quarter of 2025 alone it reported a net profit
of $12.8 billion USD. And in the following three months, TSMC posted net
profits of $14.76 billion.

ΓÇ£TSMCΓÇÖs secret sauce is its mastery of yield,ΓÇ¥ explained ARPU
Intelligence, an analyst group that prefers to use the group name over
individual attribution. ΓÇ£This expertise is the result of decades of
accumulated process refinement [and] a deep institutional knowledge that
cannot be replicated.ΓÇ¥ This deep institutional knowledge and ability to
deliver high quality product creates a ΓÇ£powerful technical lock-in, since
companies like Apple and NVIDIA design their chips specifically for TSMCΓÇÖs
unique manufacturing process … It’s not as simple as sending the [chip]
design to another factory,ΓÇ¥ it added.

The downside, at least for the wider technology industry, is that TSMC is now
a bottleneck that the whole industry has come to rely upon. In the
companyΓÇÖs most recent financials, it said more than three quarters of its
business comes from North American customers. And in a call with investors,
Chairman and CEO C.C. Wei talked about the efforts the company has made to
narrow the gap between the enormous demand and its constrained supply. While
he was reticent to be specific, he did say that the companyΓÇÖs capacity is
ΓÇ£very tight,ΓÇ¥ and would likely remain that way for the foreseeable
future.

In fact, TSMCΓÇÖs capacity is so tight that itΓÇÖs already caused at least
one major name a significant headache. Earlier this year, Reuters reported
that NVIDIA canceled an order of its H20 AI chips after being informed the US
would not permit them to be exported to China. Once the ban was lifted,
however, NVIDIA was unable to find space in TSMCΓÇÖs schedule, with the next
available slot at least nine months later.

ΓÇ£TSMC has no room for error,ΓÇ¥ said ARPU Intelligence, ΓÇ£any minor
disruption can halt production with no spare capacity to absorb the shock.ΓÇ¥
It cited the Hualien earthquake which struck Taiwan on April 3, 2024, and how
it negatively impacted the number of wafers in production.

Naturally, TSMC is spending big to increase its production capacity for its
customers, both in Taiwan and the US. Close to its home, construction on its
A14 fab is expected to begin in the very near future, with the first chips
due to be produced in 2028. That facility will harness TSMCΓÇÖs A14 process
node, producing 1.4 nm chips, which offer a speed boost over the 2nm silicon
that's expected to arrive in consumer devices next year.

Image of TSMC's Arizona CampusTaiwan Semiconductor Manufacturing Co. Ltd.

Meanwhile, work continues apace on building out TSMCΓÇÖs sprawling facility
in Arizona, which broke ground in April 2021. As Reuters reported at the
time, the first facility started operating in early 2025, producing 4 nm
chips. Last week, NVIDIA and TSMC showed off the first Blackwell wafer
produced at the Arizona plant ahead of domestic volume production.

Plans for the operation have grown over time, expanding from three facilities
up to six to be built over the next decade. And while the initial outline
called for the US facilities to remain several process generations behind
Taiwan, that is also changing. In his recent investors call, Chairman and CEO
C.C. Wei pledged to invest more in the US facility to bring it only one
generation behind the Taiwanese facility.

No amount of investment from TSMC or catch-up from rivals like Samsung and
Intel will solve the current bottleneck swiftly. It will take many years, if
not decades, for the world to reduce its reliance on Taiwan for bleeding-edge
manufacturing. TSMC's island remains the industry's weak point, and should
something go wrong, the consequences could be dire indeed.

This article originally appeared on Engadget at
https://www.engadget.com/computing/openais-re...
pressure-on-tsmc-130000194.html?src=rss

---
VRSS v2.1.180528