OpenAI’s latest chip offers heap extra strain on TSMC

In latest weeks, OpenAI has signed blockbuster offers with AMD and Broadcom to construct huge numbers of AI chips. A lot of the main focus has been on the monetary implications, since OpenAI will want a whole bunch of billions of {dollars} to make good on its guarantees. As essential as it’s to have a look at the fairly implausible financials, we additionally want to have a look at the broader implications for the trade. Like, the chips themselves, what that spells for the AI trade as an entire, and the added strain on TSMC, the one chip firm that may really construct these things.

The Offers

OpenAI’s deal with AMD will see the chip big construct out 6 gigawatts’ (GW) value of GPUs within the subsequent few years. The primary 1 GW deployment of AMD’s Intuition MI450 silicon will begin within the again finish of 2026, with extra to return. AMD’s CFO Jean Hu believes that the partnership will ship “tens of billions of {dollars} in income” in future, justifying the difficult means the deal is funded.

In the meantime, Broadcom’s cope with OpenAI will see the pair collaborate on constructing 10 gigawatts’ value of AI accelerators and ethernet techniques that it has designed. The latter will likely be essential to hurry up connections between every particular person system in OpenAI’s deliberate information facilities. Just like the cope with AMD, the primary deployments of those techniques will start within the again half of 2026 and is ready to run by 2029.

Phil Burr is head of product at Lumai, a British firm trying to change conventional GPUs with optical processors. He’s received 30 years expertise within the chip world, together with a stint as a senior director at ARM. Burr defined the nitty-gritty of OpenAI’s offers with each Broadcom and AMD, and what each imply for the broader world.

Burr first poured water on OpenAI’s declare that it could be “designing” the gear produced by Broadcom. “Broadcom has a large portfolio of IP blocks and pre-designed components of a chip,” he mentioned, “it should put these collectively in keeping with the specification of the client.” He went on to say that Broadcom will basically put collectively a collection of blocks it has already designed to swimsuit the specification laid down by a buyer, on this case OpenAI.

Equally, the AI accelerators Broadcom will construct are geared towards extra environment friendly working of fashions OpenAI has already skilled and constructed — a course of referred to as inference in AI circles. “It might probably tailor the workload and cut back energy, or improve efficiency,” mentioned Burr, however these advantages would solely work in OpenAI’s favor, quite than for the broader AI trade.

I requested Burr why each firm within the AI area talks about gigawatts value of chips quite than in additional easy numbers. He defined that, usually, it’s as a result of each events don’t but know what number of chips could be required to satisfy these lofty objectives. However you may make an inexpensive guess if you happen to knew the ability draw of a particular chip divided by the general objective, then lower that quantity in half, then take away an additional 10 p.c. “For each watt of energy you burn within the chip, you want a couple of watt of energy to chill it as effectively.”

When it comes to what OpenAI will get from these offers, Burr believes that the startup will get monetary savings on chips, since there’s “much less margin” from making your personal versus shopping for gear from NVIDIA. Plus, having the ability to produce customized silicon to tailor the work to their wants ought to see vital velocity and efficiency positive aspects on rival techniques. After all, the subsequent greatest profit is that OpenAI now has “variety in provide,” quite than being reliant on one supplier for all its wants. “No person desires a single provider,” mentioned Burr.

The Manufacturing facility

Besides, after all, OpenAI could also be sourcing chips from quite a lot of its companions, however it doesn’t matter what’s stamped on the silicon, all of it comes from the identical place. “I’d be very shocked if it wasn’t TSMC,” mentioned Burr, “I’m fairly positive all the AI chips on the market use TSMC.” TSMC is brief for Taiwan Semiconductor Manufacturing Firm which, over the past decade, has blown previous its main rivals to turn into the largest (and in lots of instances solely) supply of bleeding-edge chips for the entire expertise trade. In contrast to historic rivals, which designed and manufactured their very own {hardware}, TSMC is a pure play foundry, solely constructing chips designed by others.

Gil Luria is Managing Director at head of expertise analysis at funding agency DA Davidson. He mentioned that TSMC isn’t only a bottleneck for the western expertise trade, however in truth is the “best single level of failure for your complete international economic system.” Luria credit the corporate with a powerful enlargement “contemplating it has needed to ramp the manufacturing of GPUs tenfold over the past three years.” However mentioned that, “in a catastrophic situation the place TSMC just isn’t capable of produce in Taiwan, the disruption could be vital.” And that gained’t simply have an effect on the AI world, however “cellular handset gross sales in addition to international automobile gross sales.”

TSMC supplanted Intel for numerous well-documented reasons, however probably the most related right here is its embrace of Excessive Ultraviolet Lithography (EUV). It’s a expertise that Intel had initially backed, however struggled to completely undertake, permitting TSMC to choose it up and run straight to the highest. EUV produces the headline-grabbing chips utilized by just about everybody within the shopper electronics world. Apple, Qualcomm, NVIDIA, AMD (together with the SOCs contained in the PS5 and Xbox) all use TSMC chips. Even Intel has been utilizing TSMC foundries for some shopper CPUs because it races to bridge to gulf in manufacturing between the 2 corporations.

“TSMC is the present chief in superior 3 nanometer (nm) course of applied sciences,” mentioned College of Pennsylvania Professor Benjamin C. Lee. The corporate’s solely significant rivals are Intel and Samsung, neither of which pose a risk to its dominance at current. “Intel has been working for a really very long time to construct a foundry enterprise,” he defined, “however has but to good its interface.” Samsung is in the same state of affairs, however Professor Lee defined it “has been unable to draw sufficient clients to generate a worthwhile manufacturing enterprise.”

Professor Lee mentioned that TSMC, by comparability, has turn into so profitable due to how good its chips are, and the way straightforward it’s for purchasers to construct chips with its instruments. “TSMC fabricates chips with excessive yield, which is to say extra of its chips emerge from the fabrication course of at anticipated efficiency and reliability.” Consequently, it must be no shock that TSMC is a cash making machine. Within the second quarter of 2025 alone it reported a net profit of $12.8 billion USD. And within the following three months, TSMC posted net profits of $14.76 billion.

“TSMC’s secret sauce is its mastery of yield,” defined ARPU Intelligence, an analyst group that prefers to make use of the group title over particular person attribution. “This experience is the results of a long time of amassed course of refinement [and] a deep institutional data that can’t be replicated.” This deep institutional data and skill to ship prime quality product creates a “highly effective technical lock-in, since corporations like Apple and NVIDIA design their chips particularly for TSMC’s distinctive manufacturing course of … It’s not so simple as sending the [chip] design to a different manufacturing facility,” it added.

The draw back, a minimum of for the broader expertise trade, is that TSMC is now a bottleneck that the entire trade has come to depend on. Within the firm’s most up-to-date financials, it mentioned greater than three quarters of its enterprise comes from North American clients. And in a name with investors, Chairman and CEO C.C. Wei talked concerning the efforts the corporate has made to slim the hole between the big demand and its constrained provide. Whereas he was reticent to be particular, he did say that the corporate’s capability is “very tight,” and would doubtless stay that means for the foreseeable future.

In actual fact, TSMC’s capability is so tight that it’s already precipitated a minimum of one main title a big headache. Earlier this 12 months, Reuters reported that NVIDIA canceled an order of its H20 AI chips after being knowledgeable the US wouldn’t allow them to be exported to China. As soon as the ban was lifted, nonetheless, NVIDIA was unable to search out area in TSMC’s schedule, with the subsequent accessible slot a minimum of 9 months later.

“TSMC has no room for error,” mentioned ARPU Intelligence, “any minor disruption can halt manufacturing with no spare capability to soak up the shock.” It cited the Hualien earthquake which struck Taiwan on April 3, 2024, and the way it negatively impacted the variety of wafers in manufacturing.

Naturally, TSMC is spending large to extend its manufacturing capability for its clients, each in Taiwan and the US. Near its residence, development on its A14 fab is expected to begin in the very near future, with the primary chips attributable to be produced in 2028. That facility will harness TSMC’s A14 course of node, producing 1.4 nm chips, which provide a velocity increase over the 2nm silicon that is anticipated to reach in shopper units subsequent 12 months.

In the meantime, work continues apace on constructing out TSMC’s sprawling facility in Arizona, which broke floor in April 2021. As Reuters reported on the time, the primary facility began working in early 2025, producing 4 nm chips. Final week, NVIDIA and TSMC confirmed off the primary Blackwell wafer produced on the Arizona plant forward of home quantity manufacturing.

Plans for the operation have grown over time, increasing from three amenities as much as six to be constructed over the subsequent decade. And whereas the preliminary define referred to as for the US amenities to stay a number of course of generations behind Taiwan, that can also be altering. In his recent investors call, Chairman and CEO C.C. Wei pledged to speculate extra within the US facility to deliver it just one era behind the Taiwanese facility.

No quantity of funding from TSMC or catch-up from rivals like Samsung and Intel will resolve the present bottleneck swiftly. It can take a few years, if not a long time, for the world to cut back its reliance on Taiwan for bleeding-edge manufacturing. TSMC’s island stays the trade’s weak level, and may one thing go incorrect, the implications could possibly be dire certainly.