Chips and hardware, where it stands now

If you run artificial intelligence (AI) in a pharma or biotech company, if your team uses a cloud AI service, the chips sit with your provider. The chain runs from the chip makers, to the cloud or AI vendor that buys their chips, to your team using that vendor’s service. How many chips exist, and who may buy them, shapes how much capacity those vendors can offer, how soon, and on which hardware.

In my reading, chip news reaches your team indirectly, through your vendors, and most of it never will. Where it does, it shows up in the compute budget, vendor risk, change control on GxP adjacent systems (those touching regulated good practice work), and governance. In 2026 memory is short, advanced packaging is tight, most chips used in AI data centers are made in Taiwan and South Korea, and export rules keep changing.

This is the second piece in a nine-part series under Where each part of the AI pipeline stands now. The whole chain is mapped in The nine parts of the AI pipeline. Chips use the electricity from part 1, energy. Most figures come from International Energy Agency (IEA) reports of April 2025 and April 2026, updated where newer figures exist, such as late September 2026 memory prices. I flag company claims as such.

Why GPUs and purpose-built chips do the work

Laptops run on a central processing unit (CPU), a general purpose chip. Models need enormous numbers of simple calculations that can run at once, which a graphics processing unit (GPU) is designed for. The IEA’s April 2025 Energy and AI report says GPUs currently dominate AI computing because that design makes training and running models much faster and more energy efficient.

The IEA calls GPUs and chips designed for particular AI jobs accelerators, such as Google’s tensor processing unit or an application-specific integrated circuit built for one job, like running trained models. The IEA’s Key Questions on Energy and AI, published April 16, 2026, says these task-specific chips far outperform CPUs on machine learning, and for some targeted jobs get more done per watt than current AI GPUs.

I would ask each vendor whether your workload can move to another accelerator, and whether you will be told. For GxP systems, I treat new hardware or a new cloud region as a change that calls for an impact review, even if the model name stays the same. The review decides what, if anything, needs retesting, scaled to the context of use, meaning the system’s intended job and conditions (see How much validation does your AI need).

The chip is one part of a larger machine

The April 2026 report says chips are now networked so tightly that many work as a single computer. Four things sit around the chip.

• MemoryMemory close to the chip saves time and energy. Advanced GPUs need large amounts of high-bandwidth memory (HBM), fast memory placed beside the processor. Citing the research firm SemiAnalysis, the IEA says leading GPU systems held under 150 gigabytes (GB) of memory in 2023 and will hold over 1,000 GB by 2027.
• PackagingPackaging, the last stage of making a chip, encloses the bare silicon in a protective case that brings power in and carries heat out. The April 2026 IEA report says packaging capacity for high-end chips was a supply chain bottleneck in 2025.
• NetworkingNvidia's [January 5, 2026 release on Rubin](https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer) says each Rubin GPU can link to the others at 3.6 terabytes a second. That is Nvidia's own figure.
• RacksThe same release describes the Vera Rubin NVL72 rack: 72 Rubin GPUs and 36 of Nvidia's Vera CPUs in one cabinet. Nvidia said in January that Rubin chips were in full production, with partners' systems due in the second half of 2026.

The chart shows power per rack, counting processors, networking and storage too. The IEA still treats Rubin as announced, so I show it that way. With Rubin, the report says, a fridge sized box would draw at peak as much as about 65 households, and its summary says a rack could reach that by 2027. The energy explainer covers cooling.

For pharma teams, the lag from announcement to use matters. When a roadmap names a new chip, I would ask when it reaches your data’s regions, and whether tested workloads stay put until you move them.

What a chip is made of

An AI chip package has six main parts. Some materials stay in the finished chip; others, such as gases and light sensitive coatings, are only used to make it.

1. Transistors. The billions of switches that calculate, formed by printing circuit shapes with light, then etching and depositing material, over and over. Key materials: silicon and silicon germanium in the chip; photoresists (light sensitive coatings, washed away afterward) and neon, a laser gas in older deep ultraviolet printing machines. Supply now: on December 3, 2025, TrendForce reported, citing Commercial Times, rumors that Japan had tightened photoresist exports to China. The same report said “Tokyo has issued no official export ban.” It put Japan’s share of the photoresist market above 70%. The sources here say nothing about current neon supply.

2. Wiring layers. Layers of fine metal lines above the transistors connect them and carry power. Key materials: copper for the lines, with cobalt in some contacts and liners. Supply now: the sources here say nothing about copper or cobalt supply for chipmaking.

3. The die. One piece of processed silicon cut from a round wafer, usually sliced from a crystal grown from molten silicon in a high purity quartz crucible. Key materials: silicon wafers and quartz, plus helium to cool the factory machines, covered in the memory section. Supply now: on July 18, 2026, RFI reported that the Spruce Pine district in North Carolina is “estimated to supply more than 80 percent of the world’s ultra-pure quartz.” The Quartz Corp had said in June it was closing its plant there because of losses in the solar industry, while Sibelco keeps expanding. That is a concentration risk, and no shortage was reported. China’s limits on gallium and germanium are in Export controls.

4. Packaging. Joins the compute dies and memory stacks into one unit for mounting on a board. Key materials: copper bumps and pillars, solder and molding compounds. Supply now: on June 15, 2026, TrendForce reported that institutional investors, cited by Economic Daily News, expected the gap between supply and demand for TSMC’s CoWoS packaging to narrow from about 20% to about 10% by the end of 2026. In August, a TSMC vice president for advanced packaging said capacity was close to meeting demand.

5. HBM memory stacks. Several memory dies piled on a base die, joined by copper connections through the silicon and tiny solder bumps, placed beside the processor. Key materials: silicon, copper and solder. Supply now: HBM supply is tight; the dates and figures are in the memory section.

6. Interposers and substrates. The interposer, a thin wired layer, links the processor to its memory stacks; some are a full silicon slab, newer ones copper wiring with small silicon bridges. It sits on a substrate whose core is reinforced with glass fiber cloth, with copper wiring separated by layers of an insulating film called ABF. Key materials: silicon, copper, ABF film and a special glass fiber called T glass. Supply now: a March 10, 2026 summary of an earnings call by the substrate maker Kinsus put the T glass shortage at about 10 to 15%, possibly lasting into the first half of 2027. In August, the same TSMC vice president, cited by Next Apple, warned of tight ABF substrate supply over the next few years.

The reports cited above describe conditions on their publication dates.

I read this list the way I would read a raw material review: one district supplying most of a critical input, a plant closing, a shortage into 2027. None of it lands on a pharma team directly. It sits two steps back, behind the chip makers and your vendors, so I would ask vendors which of these inputs their planned capacity depends on, then record any relevant supply or concentration risk.

Performance per watt, and the link back to energy

Performance per watt is how much computing a chip does for each watt it draws. Drawing on Epoch AI’s public database of machine learning hardware, covering more than 170 accelerators, the April 2026 report says leading AI chips improved performance per watt by roughly 30% to 40% a year on average over the past decade. Gains are uneven and slowing as transistor scaling matures.

A growing share of energy goes to moving data between memory, chips and racks, the report says. Optical links, carrying signals as light with less energy, are entering production between racks, with chip to chip links expected around 2027 to 2028; the overall saving is uncertain.

I would not count on chip efficiency to shrink a pharma AI compute budget. If you report IT energy use, ask vendors which hardware generation runs your workloads.

Who designs the chips, and who makes them

Intel designs and makes chips. Others use a foundry, which manufactures to other firms’ designs, such as Taiwan Semiconductor Manufacturing Company (TSMC) or Samsung Foundry. The 2025 report calls Nvidia, Broadcom, AMD and Intel the dominant designers of AI chips. (The IEA calls Taiwan Chinese Taipei.) Nvidia’s January release lists reliance on outside packaging and testing firms as a risk.

TrendForce figures show both ends are concentrated. Among chip designers without their own factories, Nvidia’s second quarter 2026 revenue was nearly $91.2 billion, against about $18.9 billion for second place Broadcom, it reported in September. TSMC took 72.5% of foundry revenue that quarter, up from 58.5% at the end of 2022, while Samsung’s share fell from 15.8% to 5.9%.

Citing a 2021 industry study, the IEA puts more than 70% of chip production in East Asia. In a March 4, 2025 release, TSMC said it intended to raise planned U.S. investment to $165 billion, and that its first Arizona plant had been in volume production since late 2024.

In my view, for vendor risk assessment, I would check whether an alternative vendor shares the same chip suppliers and manufacturing exposure, ask each vendor how it plans for disruption in Taiwan or South Korea, and ask about continuity and recovery: where workloads would run, how long a move would take, and what you would be told. Keep a fallback for AI systems your operations depend on.

Export controls

Export controls set which goods and technology can go abroad, and to whom, without a license. The United States limits who gets advanced AI chips; China limits some minerals used to make them. U.S. rules, set by the Commerce Department’s Bureau of Industry and Security, have changed several times since early 2025. This section is dated October 15, 2026.

• May 13, 2025. In a press release, the bureau announced it would rescind the AI Diffusion Rule, issued that January, which would have sorted countries into tiers for access to advanced chips. It stopped enforcing the rule and said a replacement would follow.
• January 15, 2026. A final rule in the Federal Register changed how license requests to export chips such as Nvidia’s H200 and AMD’s MI325X from the United States to China and Macau are judged, from a presumption of denial to case by case. The exporter must show enough supply at home and no loss of factory capacity for U.S. buyers, the buyer must screen its customers and secure the chips, and an independent U.S. lab must check each shipment before it leaves. Combined exports to China and Macau are capped at half the computing power of the same chips already shipped for U.S. use. More powerful chips, and shipments onward from other countries, still face a presumption of denial, meaning applications are normally refused.
• March 13, 2026. Reuters reported that Commerce had withdrawn a draft replacement for the AI Diffusion Rule from review by other agencies. I found no replacement rule in the Federal Register up to October 15, 2026.
• May 31, 2026. In guidance, the bureau said a license is still needed, unless an exception applies, to send advanced AI chips, or equipment containing them, to any company headquartered in China or Macau, or whose ultimate parent company is, even when the company’s site is in another country.
• July 10, 2026, published July 14. A rule moved the United Arab Emirates into a more favored group of countries, opening license-free routes when the government agencies or approved companies receiving and using the chips meet its conditions. Other buyers there still need licenses.

A rule extending the restrictions to companies at least half owned by blacklisted firms was suspended for a year in November 2025. The suspension is set to end on November 9, 2026, unless it is extended.

Inside China, the April 2026 report says, model developers are short of high-end chips, and China’s own AI chips lag behind leading chips from suppliers such as Nvidia in energy efficiency.

If your AI comes through vendors, I would ask how these rules affect their hardware supply. If you have sites or partners in China, or partners with a Chinese parent, ask trade compliance how the May 31 guidance applies before AI hardware goes there. For multiyear AI contracts, note the November 9 and November 27, 2026 suspension end dates and ask how a change could affect promised capacity.

The slow step now is memory

The IEA’s figures on the squeeze:

• HBM takes about three times the wafer capacity of ordinary memory for the same number of gigabytes, a figure the IEA takes from the memory maker Micron. A wafer is the thin silicon disc chips are made on.
• Memory inventory fell from almost four months in late 2024 to about three weeks in late 2025.
• Memory prices rose about tenfold between late 2024 and early 2026, the IEA says.
• A new chip factory takes about two years to come into operation.
• An analyst estimate the IEA reports puts current HBM output at enough for roughly 25 gigawatts (GW) of new AI-ready servers a year through 2027. GW here measures the servers’ combined power rating, which is separate from the electricity they use. The IEA calls that broadly in line with its main projection, leaving little room for faster growth.

TrendForce agrees. Contract prices for conventional dynamic random-access memory (DRAM), computers’ main working memory, rose about 93 to 98% in the first quarter of 2026 over the quarter before, it said on June 1, while HBM prices were declining. In March it had forecast a further DRAM rise of 58 to 63% for the second quarter. By September 30 it expected a smaller 10 to 15% rise in the fourth quarter, with the market still short of supply, and on September 29 it said memory supply would stay tight through 2027.

Helium is a second 2026 risk. It cools and cleans advanced chipmaking machines, the IEA says, and nothing works as well. The Middle East conflict damaged production in Qatar, which had supplied roughly one third of the world’s helium, and chipmakers in Taiwan and South Korea, who get more than half of theirs from the Gulf, were drawing on stockpiles.

Memory prices are one place chip supply can show up in a pharma AI budget. If you plan to buy on premises servers for 2027, I would get quotes early and check delivery dates against the plan. If you rent through a cloud or AI vendor, ask how much notice comes before regional capacity changes.

My reading for pharma and biotech teams

Memory is short, and I expect it to pace new AI servers until new factories open. A few companies in Taiwan and South Korea make most AI data center chips, and helium shows a distant conflict can slow them.

I would watch for capacity limits, changes in quotes and delays between announcements and availability. I would turn each risk into a question for every vendor behind a system you depend on.

For GxP systems, here is how I would walk a hardware or capacity change through an impact review.

Give these questions an owner and reopen them at every hardware or region notice. Add continuity and recovery questions too: what the vendor’s recovery plan covers, how fast it has been tested, and whether your validated configuration comes back the same. My AI Project Assessment Framework weighs value against lifecycle costs, including change work from capacity shifts. For GxP adjacent systems, I would log the model version per call, as in my ALCOA+ audit trail for LLM calls, and link each vendor hardware or region notice to its change record. The rest of my tools are on the frameworks page, and if you want help turning these questions into a vendor review, see Work with me.

What’s happening now

This page is kept current. Short, dated Right now posts on chips:

• Data centers in space: SpaceXAI and Nvidia’s plan, announced on August 24, 2026, for a Starmind AI satellite built on an adapted version of the Vera Rubin rack system described above.

The short account of all nine parts, with links as they are written, is Where each part of the AI pipeline stands now.