Ever since Nvidia became one of the most valuable companies on the planet, we've seen a few different versions of Jensen Huang.
There's the leather-jacket charmer. There's food-blogger Jensen, working his way through street-food stalls. And in late May, we got the rare one: dancing Jensen. (Like Pokémon cards, you gotta catch 'em all!)
At the launch party for Nvidia's new Taiwan headquarters in Taipei, the most important man in the AI supply chain got up on stage and bust out some moves.
Great moves for a 63-year-old! Looking like anything but a trillion-dollar CEO. Jensen Huang at Nvidia Event in Taipei (Source: Moneycontrol)
That same week, a few miles away on the Taiwan Stock Exchange, something quietly historic happened: Taiwan overtook India and the UK to become the fifth-largest stock market in the world.
An island of 23 million people, fewer than in greater Delhi, now hosts a bigger equity market than all of India, $4.95 trillion to $4.92 trillion, behind only the US, mainland China, Japan, and Hong Kong.
And yes, the obvious caveat is that more than 40% of that market is a single company. TSMC, the largest chipmaker on earth, is up nearly 50% this year and has dragged the whole index to record highs behind it.
So the lazy read would be that Taiwan is just TSMC with extra steps. But that read is wrong and it gets more wrong every quarter.
Standing in front of his employees that same week, Huang called Taiwan the "epicenter of the AI revolution" and not only for chips. Taiwan, he said, is where the chips, the advanced packaging, the systems, and the AI supercomputers themselves get made. The whole stack.
TSMC is the headline act. It is not the whole show.
This piece is about the rest of the show. How deep Taiwan's AI ecosystem really runs, why it's becoming indispensable to the buildout, and a few under-watched names worth a closer look.
What The Island Actually Makes
Taiwan’s story is a full-stack hardware story that happens to have a chip company sitting on top.
Start with the foundry, because it earns the headline. TSMC fabricates ~72% of the world's leading-edge logic, effectively every frontier AI accelerator, Nvidia's GPUs and the hyperscalers' custom silicon alike. But the chip is only the first station on a long Taiwanese assembly line.
That chip has to be packaged. TSMC's CoWoS lines and ASE's OSAT plants are the global chokepoint for fusing GPU and memory into one module. It has to sit on an advanced substrate, a near-duopoly of Taiwanese and Japanese names. It has to be drilled into ultra-dense boards, cooled, powered, wired with optics, and bolted into a rack, which Taiwanese ODMs (Foxconn, Quanta, Wiwynn, Wistron, Inventec) then assemble, accounting for the overwhelming majority of the world's AI servers.
Nearly every one of those steps happens in Taiwan or in Taiwanese-owned plants for Nvidia and many other AI processor companies. When Huang says each Vera Rubin system runs on roughly 150 ecosystem partners and close to 2 million parts, that's what he means: the island isn't merely a supplier to the AI buildout but it is buildout's factory floor.
And that floor is running hot.
At GTC Taipei, Huang said the Vera Rubin supply chain is twice the size of Grace Blackwell's and that a rack which took two hours to assemble now takes five minutes. Capacity up, throughput up, all of it straining to keep pace.
Here's how you know it isn't just talk. Taiwanese companies report revenue monthly, by the 10th of the following month, a disclosure cadence almost no other major market requires. On Tessara Terminal, we have a curated list of about 35 Taiwan companies across various domains of the AI supply chain like foundry, packaging, cooling, etc.
Of these 35 companies, 19 hit all-time-high monthly revenue between March and May 2026. Substrate, drills, cooling, controllers, ASICs, ODMs, all setting records at once.
So the larger demand in AI is driving the the entire ecosystem up.
The Show Isn't Over
A market up more than 50% on the year invites the obvious question: is the move done? The demand signals say no and many of the tailwinds for this ecosystem are structural. Some of them include:
The chip giants are ramping their Taiwan spend, fast.
Jensen put hard numbers on it in May when he mentioned that Nvidia's annual spend in Taiwan has gone from $10–15 billion five years ago to around $100 billion today, on a path to ~$150 billion. A tenfold jump in five years.
And it isn't only Nvidia. AMD committed more than $10 billion across Taiwan's chip ecosystem this year, including a deal with ASE to scale the 2.5D packaging behind its next-generation EPYC line of CPUs.
That combined spend from these two chip companies and a few others is the ecosystem's order book. Every dollar lands on a Taiwanese foundry, packaging line, substrate, cold plate, or ODM. And with multiple buyers now competing for the same constrained lines, those suppliers get pricing power on top of volume.
Hyperscaler capex is heading toward $1 trillion a year.
The big 5 hyperscalers (Google, Amazon, Microsoft, Meta, Oracle) are guiding to roughly $725 billion of combined capex in 2026, up 77% year-on-year, and many analysts (us included) now expect that number to hit $1 trillion in 2027.
Yes, a growing share of that goes to other aspects of the data center buildout like power, land, and buildings rather than chips. But the accelerator slice still rises in absolute dollars every year and almost every accelerator gets fabricated, packaged, and racked in Taiwan. A bigger total pie means a bigger Taiwan slice.
Every major hyperscaler is now building its own AI chip and almost all of it runs through Taiwan.
Google (TPUs), Amazon (Trainium), Microsoft (Maia 200), and Meta (MTIA) all have custom accelerators in production, and OpenAI is now designing its own with Broadcom.
Custom-ASIC shipments are projected to grow ~45% in 2026 against ~16% for merchant GPUs. The instinct is to read this as a threat to the "Nvidia trade" but for Taiwan it's the opposite.
TSMC fabricates roughly 92% of advanced AI chips and makes silicon for all five hyperscalers. Maia 200 and Trainium 3 are built on TSMC 3nm, and Meta's top-end MTIA uses TSMC CoWoS packaging. The design work increasingly flows to Taiwanese houses too. MediaTek, Alchip, and TSMC's own GUC are big beneficiaries. Custom silicon demand routes straight through Taiwan, adding a second demand engine alongside the GPU.
Taiwan's order book is already at a record.
Export orders, which is the book that precedes the shipment, hit a record $91.1 billion in March, up 65.9% year-on-year, the fastest growth in sixteen years and ~20 points above expectations.
Orders for information-and-communications products grew 120.9%, nearly double the headline. Because orders lead actual revenue by months, this is the ecosystem's pipeline filling up in advance. The next few quarters of Taiwan-eco growth are largely already booked, not forecast.
Each new generation of packs more Taiwan content into every rack.
The jump from Grace Blackwell to Vera Rubin re-architected the box. Liquid cooling deepened, high-voltage DC power and co-packaged optics came in, and the part count climbed across the board. Nvidia says the Rubin supply chain is twice the size of Blackwell's and an independent bill-of-materials estimate from Morgan Stanley lands almost exactly there.
A single VR200 NVL72 rack stands at ~$7.8 million against ~$4.0 million for GB300. That's +95%, near enough to double.
The story is where that extra $3.8 million goes. Break the BOM apart and almost every line that grows is one Taiwan touches. From denser boards drilled by Topoint and the PCB houses (Gold Circuit, Unimicron), substrate from Nan Ya and Kinsus, power from Delta, cooling from Auras and AVC, all racked by ODMs like Wiwynn, Quanta, and Foxconn. PCB content alone more than triples per rack, +233%.
Obviously, not all of that jump is more content per rack. Part of it is shortage pricing with substrates, drills, and memory are all sold out, and tight supply lifts the dollar figure on its own. But part is real content growth, more components designed into every Rubin rack. Either way, the dollars land on the same Taiwanese suppliers.
A Few Names Worth Watching
So the demand is real, and it's broad. The natural next question is who plays it. Beyond the giants everyone already owns, the more interesting answer sits further down the stack, where the edge comes from owning the scarce input.
For each name, we lead with the constraint it sits in, like substrates, storage, cooling, optics, etc., and how tight that constraint is right now. On Tessara, we score every constraint by how tight it is running.
Mcap: $15.9B | Fwd P.E: 8.5 | Constraint: NAND Flash
What it does:
Phison designs the NAND flash controllers and enterprise SSDs that AI inference and agents lean on to hold state, cache context, and move data. This is the other place value leaks past the GPU. The controller is the part that makes raw flash usable. It handles error correction, wear-leveling, and the traffic between memory and host. Every NAND chip that ships needs one. That makes it a quiet tollgate on the entire storage layer and storage demand can't grow without controllers growing alongside it.
Phison sits at the center of that layer, with roughly a fifth of the SSD-controller market and a fast-growing enterprise SSD business.
Why it matters:
The shortage backdrop is unusually strong. TLC NAND prices roughly doubled in six months, and Phison’s CEO has warned that the imbalance could last for years, with most 2026 capacity already sold out.
Phison locked in supply early and is steering more of it toward higher-margin enterprise and industrial customers. The numbers show it with March revenue hitting a record NT$18.3 billion, up 221% year-on-year, and enterprise SSDs rising to around 30% of Q1 revenue.
What to watch:
The forward angle is inference. As AI shifts from training big models to running them constantly, storage starts to matter more. Phison’s aiDAPTIV+ is built for that world. In simple terms, it lets AI systems use flash storage as an extra memory layer when GPU memory and system RAM are not enough.
That matters for long-context inference, agentic workflows, and KV-cache reuse, where models need to remember and reuse more information. Phison says this can cut inference token costs by more than 70%.
The ecosystem is also gaining credibility, with demos or deployments tied to Nvidia DGX Spark, Supermicro, MSI, Acer, and RedData. Gen5 Pascari enterprise drives are already shipping, and newer controllers such as E28 reflect Phison’s push to make storage more AI-aware. Management believes 2026 will mark a broader shift toward inference and Phison wants storage to become part of that AI memory stack.
The catch:
This is still an emerging AI angle layered on top of a cyclical NAND business. NAND shortages help pricing today, but they can reverse. And aiDAPTIV+ is more of an on-prem, edge, and enterprise inference opportunity for now, not yet a proven hyperscaler standard. The thesis is interesting because Phison has both near-term shortage leverage and a real option on AI inference storage. But it is still a storage company first.
Mcap: $92B | Fwd P.E: 12.6 | YTD: +16% | Constraint: Taiwan Sever ODMs
What it does:
Wiwynn is the ODM that integrates compute, cooling, power, networking/optics, and rack manufacturing into deployable systems. Its two biggest customers are the hyperscalers themselves. Meta accounts for more than half of revenue, and Microsoft is the other anchor, which is both the strength and the concentration risk.
What to watch:
Q1 2026 was a record, with consolidated revenue of NT$276.5 billion, up 62% year-on-year, with profit after tax up 44%. Wiwynn (with Wistron) is among the first ODMs ready for Vera Rubin NVL72. Fully liquid-cooled, with HVDC power and co-packaged-optics designs in hand. And it's bringing AMD's Instinct-based Helios rack into production, so the franchise isn't a pure Nvidia bet.
Looking forward, it has tripled capex to fund a new $300 million Texas plant (running since end-2025), a Kaohsiung site, and expansions in Malaysia and the Czech Republic, and it has pre-secured power for high-density rack integration through 2028, capacity being the thing that actually gates this business.
Two smart moves: First, Wiwynn invested in Ayar Labs, the co-packaged-optics leader, pushing the rack builder higher up the stack as optical interconnects become more important for next-gen AI systems.
Second, it changed how it handles memory for certain customers. Instead of buying DRAM/NAND and booking it as revenue, Wiwynn now acts more like a procurement agent, so that memory is excluded from reported revenue. Because memory is scarce, expensive, and lower-margin, this makes headline revenue look softer but should improve the quality of margins. So the better way to judge Wiwynn from here is shipments, gross margin, and profit.
Wiwynn had a record breaking Q1 revenue (Source: Tessara)
Mcap: $2.9B | P.E: 31 | YTD: +1% | Constraint: Direct Liquid Cooling
What it does:
Auras makes the cold plates, manifolds, and rack-level thermal modules that liquid-cool an AI rack.
Why it matters:
As AI racks get denser, air cooling stops being enough and more of the rack has to move to liquid. That pushes demand away from basic heat sinks and into Auras’s lane. Auras also has exposure to Nvidia-platform servers through cooling designs for DGX, MGX, and HGX systems, but the cleaner thesis is not that Nvidia is directly buying everything from Auras. It is that Auras is selling into the broader Nvidia server ecosystem.
What to watch:
Auras raised its 2026 revenue-growth target from 50% to 70%, guided to quarterly growth through the year, and said cooling-demand visibility now stretches into 2028. Liquid cooling has gone from 12% of revenue in 2024 to more than 40% in 2025, and reportedly crossed 55% in Q1 2026.
Auras Technology revenue up 94% YoY (Source: Tessara)
The next leg is not just cold plates. Nvidia reportedly has four qualified cold-plate suppliers, while Auras is leaning more into higher-value parts like quick disconnectors, manifolds, CDUs, and rack-level cooling modules. That may be the better angle because it would mean that more liquid-cooling content per rack, even if the exact cold-plate share changes by platform.
The catch:
Cooling is the most crowded lane in this report. Auras competes with AVC, Jentech, Cooler Master, Boyd, CoolIT, and others. And because the cooling layout is ultimately decided by Nvidia, hyperscalers, and server OEMs, Auras’s content per rack can change if the design changes. That is the core tension in the thermal trade. The demand is real, but your share of the rack is still decided by someone else.
Mcap: $24B | Fwd P.E: 51 | YTD: +395% | Constraint: FC-BGA Substrate
What it does:
Nan Ya makes FC-BGA substrates, which are the dense little circuit board a GPU sits directly on top of. A modern chip has thousands of connections packed far too tightly to wire straight into a motherboard, so this board goes in between and fans them out to something the board can plug into.
The name just describes the build: the chip is flipped face-down onto the top (flip-chip), and a grid of solder balls underneath (ball-grid array) connects it down to the motherboard. The high-end ones stack 14–20 wiring layers with lines only microns wide.
These substrates are built up from ABF (Ajinomoto Build-up Film), the thin insulator laid between each copper layer. This is why they're often just called "ABF substrates." ABF is the ingredient and the FC-BGA substrate is the finished part Nan Ya ships.
Nan Ya PCB is a majority-owned subsidiary of Nan Ya Plastics within Formosa Plastics Group, and its key advantage is unusual vertical integration at the group level. Related Nan Ya entities produce glass fabric, copper foil, epoxy resin and copper-clad laminates, giving Nan Ya PCB secure access to several core substrate inputs. The one big exception is the ABF film, which it buys from Ajinomoto like everyone else.
It builds substrates at plants in Taiwan, China, and Vietnam, and one of the world’s leading advanced substrate makers, though still smaller than Unimicron.
Why it matters:
Substrates are a genuine bottleneck at critical levels right now. They are sold out across the three Taiwanese makers (Unimicron, Kinsus, Nan Ya) and underwritten into the second half of 2028. And demand climbs with every GPU generation. Blackwell needs more than double Hopper's substrate area and Rubin adds another 75% on top.
Nan Ya is the lower-profile, cheaper way to play that as it is smaller than market leader Unimicron and still mid-shift from commodity PCB maker to advanced-substrate supplier, which is exactly where the re-rating leverage sits.
We think the market is starting to re-price it as an AI name rather than a PCB house. (The whole chain also rests on one upstream chokepoint, Ajinomoto's build-up film at >95% market share, which just pushed a 30% price hike for Q3 2026, a measure of how tight things are. We covered this in our Chokepoint newsletter)
What to watch:
March revenue hit NT$4.29 billion, up 39% year-on-year (a 36-month high) as Nan Ya pushes high-end substrate for GPUs, switches, and edge AI, with capex guided toward a record in 2026. The expectation is for a double-digit growth this year as two new ASIC customers ramp into volume at once.
Nan Ya PCB up over 260% YTD
Mcap: $209B | Fwd P.E: 44 | YTD: +188% | Constraint: Custom ASICs
What it does:
MediaTek is best known as one of the world's largest chip designers, especially for the silicon behind a huge share of the world's smartphones, TVs, and Wi-Fi.
But the reason it's on this list is a newer, much smaller business of designing custom AI accelerators for hyperscalers, turning a cloud provider's chip spec into manufacturable silicon. It's the work Broadcom and Marvell have long owned.
As more AI workloads shift from pure training toward inference and custom silicon, more dollars move onto chips the hyperscalers design themselves and MediaTek wants a share of them.
Why it matters:
The key customer is reportedly Google - huge if they’re able to be a key supplier to the tech giant. MediaTek is tied to the inference side of Google’s TPU roadmap, while Broadcom remains more exposed to the higher-end training chips. That split suggests that Google is deliberately spreading its silicon work across partners, giving MediaTek a real multi-year opening in AI data centers.
Custom ASIC is the binding constraint for MediaTek (Source: Tessara)
What to watch:
MediaTek’s data-center business was tiny in 2024, but management now expects around $2 billion of revenue from the segment in 2026 and “multiple billions” in 2027. The bigger ambition is to win 10–15% of a $70–80 billion ASIC market by 2027.
The real gate is not demand, but packaging. If MediaTek cannot secure enough advanced packaging capacity from TSMC, especially CoWoS, the Google opportunity cannot fully convert into shipments. That is why its support for both TSMC CoWoS and Intel EMIB is worth watching. It is also quietly moving up the stack through a $90 million investment in Ayar Labs, the same co-packaged optics name Wiwynn backed.
The catch:
MediaTek is not a clean AI pure-play. Its core mobile business is soft, Q1 profit fell year-on-year, and the stock has already rerated hard on the custom-silicon story. The narrative is becoming real, but it is no longer early or cheap.
The bullish scenario we’re watching is that MediaTek becomes the second major custom-AI ASIC supplier after Broadcom. And we think there’s a reasonable chance this happens.
Mcap: $2.5B | Fwd P.E: 51 | YTD: +190% | Constraint: High Speed PCBs
What it does:
Topoint makes the tungsten-carbide drill bits, and the drilling services around them, that punch the holes in AI server circuit boards.
Why it matters:
As AI boards move to harder, higher-layer materials, the drills are wearing out faster. So each board burns through more bits, and the bits have to be better. At the very high end required for AI, only two companies in the world can make them: Topoint and Japan's Union Tool, and Topoint is steadily closing the share gap.
What to watch:
Demand is running so far ahead that utilization sits above 90% and the shortage is expected to persist through 2026 even after recent expansions. A new Thailand plant is ramping, and Topoint has pulled PCB heavyweights in as strategic investors.
With tungsten-carbide costs rising, Topoint pushed through a second price increase in a single quarter, while targeting a 55% high-end sales share for 2026. A commodity supplier can't raise prices like that. A bottleneck can.
Topoint is up over 185% YTD (Source:Tessara)
Mcap: $5.4B | Fwd P.E: 52 | YTD: +128% | Constraint: III-V Epi Foundry
What it does:
WIN is a III-V compound-semiconductor foundry. It fabricates the indium-phosphide lasers that carry data between racks as light, once copper runs out of room.
Why it matters:
Co-packaged optics is the next interconnect frontier. Nvidia's Rubin Ultra is designed from the ground up to integrate silicon photonics, and the broader silicon-photonics market is set to roughly triple toward $7 billion by 2031. It runs on InP lasers, which is a constrained supply chain with sub-30% wafer yields and only a handful of players who can fab at scale.
How tight? In April, Nvidia put $4 billion into Lumentum and Coherent purely to lock in laser capacity. Those two own their fabs but not every photonics company will, and that's the opening.
Sweden's Sivers Semiconductors ($SIVE), a fabless InP specialist, picked WIN as its outsourced manufacturing partner to scale up production of DFB lasers - the precise, stable lasers used to send data through fiber. As photonics demand explodes, outsourced III-V capacity becomes the scarce input, and WIN is one of the few foundries that can supply it.
The catch:
Of the names here, WIN is the earliest-stage bet. Co-packaged optics is mostly a 2027–2028 story, and WIN's revenue is still recovering off a soft base in its older business of making radio-frequency chips (the parts that handle wireless signals in phones and basestations where it holds a 70% share). We’d watch it for the option on that ramp, and on being the neutral foundry every fabless laser startup needs.
WIN Semiconductor price up over 195% YTD
A few risks to watch
So the demand is real and the companies are printing records. That's the bull case, and we think it's the right one. But a clear-eyed read means holding a few risks in view alongside it.
Customer concentration: Almost every thesis here leans on Nvidia's roadmap and a handful of hyperscalers. MediaTek's ASIC ramp is, for now, essentially one customer. Auras's content per rack is hostage to someone else's design tweak. A single platform delay, spec change, or allocation shift ripples straight through the chain.
Power: Taiwan's grid is a real ceiling. The island is energy-import-dependent, its reserve margins are thin, and an AI buildout this size is a structural new load. The constraint that bites the most may end up being electricity, and that's a slow problem to fix.
Geopolitics: This is the tail that dominates all others. Concentrating $150 billion a year and 150 Rubin partners on one island is an efficiency miracle and a strategic single point of failure at once. China-Taiwan cross-strait risk needs no elaboration. 2026 already showed the milder version, with Middle East tensions disrupting supply routes, which Taiwan's exports weathered, but it also underlined how exposed the chokepoint is.
Our View: The Taiwan Supply Chain Isn't Near Its Ceiling
Forget whether the move is "over." The better question is how long the build runs. The evidence still points to a long runway.
Taiwan’s AI ecosystem is in the early-to-middle innings of a buildout that compounds on itself. Our case rests on four visible forces.
The demand is funded and already booked: Hyperscaler capex is running near $725 billion in 2026, crosses $1 trillion in 2027, and Goldman models $5.3 trillion across the four largest spenders through 2030. That spending is already showing up in order books. Substrate is sold out into late 2028. Drills and controllers are booked through 2026. Taiwan’s export orders are at record highs. The near term is booked, and the medium term is backed by spending plans that do not reverse quickly.
Every generation loads more Taiwan into the rack: Nvidia is locked to an annual cadence with Rubin this year, Rubin Ultra in 2027, Feynman in 2028. Each generation rebuilds the rack from the ground up. Power is the clearest tell. A Blackwell rack draws north of 100 kilowatts. Rubin Ultra’s Kyber rack is specced near 600 kilowatts. That is a four-to-fivefold increase in power and cooling content in roughly two years. More content per rack means more Taiwan revenue per rack, even if unit volumes flatten
New layers of the rack keep lighting up: Co-packaged optics moves inside the rack with Kyber, pulling in more lasers. Inference pulls in more storage. Agentic AI is pushing the CPU-to-GPU ratio from 1-in-8 toward 1-to-1 (which we wrote about in our earlier research piece on CPUs), with CPU cores per gigawatt rising roughly fourfold. A new compute layer, from Vera and Rosa to Graviton and Axion, is forming around the same Taiwanese substrates, boards, and ODMs.
The buyer base keeps broadening: Five years ago, this was mostly an Nvidia story. Today it is Nvidia, AMD, every hyperscaler building custom silicon, sovereign AI buyers, and neoclouds. Different roadmaps, same bottlenecks: foundry, packaging, substrate, boards, optics, power, and cooling. Demand that no longer depends on one company’s product cycle is what turns a boom into a durable base.
Put those together and the shape is clear: more racks, more content per rack, more layers of content, and more customers buying it.
19 of the 35 Taiwan names we track hit record monthly revenue between March and May. The industrial base is climbing the value chain at once, moving from boards and cooling into optics, packaging, and custom silicon.
That is why we think this still has room to run for years.
It’s the story of an island that spent four decades learning how to make the hardest parts of the supply chain, now sitting under the most important buildout of the century.
Taiwan makes the parts the AI industry cannot do without. The chips will keep changing names. The rack will keep changing shape. But the floor underneath it still runs through Taiwan, and every generation appears to run more of it.
This issue, we named Phison, Wiwynn, Auras, Nan Ya PCB, MediaTek, Topoint, and WIN Semiconductors as names sitting on tight rungs of Taiwan's AI supply chain.
In Tessara Terminal, you can pull the live tightness score on every constraint in this piece, from NAND flash to FC-BGA substrate, screen all 35 Taiwan names we track by forward P/E and monthly revenue, and see which of 400+ public names are most exposed to the AI buildout.
On the memory side of that same build, we called Micron’s last earnings at $39.5B while the Street sat at $36B. Actual: $41.5B.
This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.