If you follow $LITE and $COHR, the AI optics buildout also runs through 17 Taiwan-listed companies.
We mapped them across five layers and rated how directly each one benefits. Supply is tightest in laser materials today, while CPO parts are mostly a 2027 and 2028 revenue story.
Top tier:
LandMark (3081 TT)
VPEC (2455 TT)
EZconn (6442 TT).
LandMark and VPEC grow the InP and GaAs epitaxial wafers that laser and photodetector chips are built on. LandMark has a four-year agreement with a US customer for data center CW lasers covering 2027 to 2030. WIN Semiconductors (3105 TT) and IntelliEPI (4971 TT) sit in the same layer, with less of their revenue tied to AI optics.
Modules and laser packaging are where 2026 demand shows up first, because pluggable 800G and 1.6T modules remain the volume product. LuxNet (4979 TT) began volume production of 800G products this year and expects 1.6T NPO mass production in the second half of 2027. Elite Advanced Laser (3450 TT) mounts laser chips onto submounts for module makers.
Fiber passives are benefiting from denser data center cabling now.
EZconn makes high fiber-count trunk cables and MPO cords. High fiber-count products passed half its sales in 1H26, when gross margin reached 59.5%. FOCI (3363 TT) is developing fiber array units for CPO and has told local media it expects shipments to pick up clearly in 2027.
Optical module boards need mSAP HDI from 800G up. Zhen Ding (4958 TT) and Compeq (2313 TT) both make them. A US broker expects the mSAP market for 1.6T and faster modules to grow from $0.8B in 2026 to $2.9B in 2028.
Accton (2345 TT) builds AI data center switches and also sells optical transceivers. How much it benefits depends on AI networking’s share of its revenue.
What we’re watching closely:
-How fast 1.6T replaces 800G in 2027
-Whether new InP substrate and CW laser capacity eases pricing
-CPO switch volumes, which decide when FAU and CPO parts turn into revenue
The AI supply chain selloff looks like a valuation reset to me, not the end of AI spending.
I think upstream orders could be cut by 10% to 20%, while many Taiwanese AI supplier stocks have already fallen 50% to 65% from their highs.
Three things hit at once.
(1) U.S. data-center approvals were already slowing.
(2) New BIS rules now require licenses to ship advanced chips to Singapore, Malaysia, Thailand, Indonesia and Brazil.
(3) Renting large amounts of compute to Chinese customers now needs a license too.
The Institute for Progress estimates China will have about 8.8 million H100-equivalents of AI compute by the end of 2027, with about 2.8 million, roughly 32%, accessed remotely in data centers outside China. About 75% of that capacity sits in Southeast Asia.
I put the loss at about 3% to 5% of new AI compute demand.
Which sounds small, but delays compound it. When a project slips a quarter, customers cancel orders placed to secure capacity, and the case for price increases weakens.
The revenue side is easier to test now that Anthropic’s S-1 is public. Earlier this week I estimated 2028 capex comes in 5% to 15% below consensus, and higher rate expectations make the funding gap harder to close.
But I think 2028 capex would still run about 50% to 70% above 2026. AI demand is definitely still growing. What has changed is the multiple investors will pay, and forecasts made before the BIS rules need redoing.
I’d worry most about suppliers that relied on overflow orders.
Leaders still sold out on core products should hold up better.
Google’s reported move back to TSMC for TPU packaging is about v10, based on what i've been hearing in the supply chain.
MoneyDJ reported today that Google tested EMIB-T for part of its next-gen TPU packaging, saw yields fall short of expectations, and recently asked TSMC for help again.
m.moneydj.com/f1a.aspx?a=04b…
I think AI circuit-board suppliers in Taiwan are heading for a split in their order books with those that relied on overflow orders losing business first.
Their stocks haven’t made that distinction yet: the group is down 50% to 65% from its 2026 highs, leaders included.
As of the Nov. 19 close, Elite Material (2383 TT), the leading supplier of high-speed CCL for AI servers, was down 52% from its high. Unimicron (3037 TT), the largest ABF substrate maker, was down 58%.
Four other CCL and substrate makers I checked were down 55% to 64%. The TAIEX was down 23%, TSMC 19%.
Many second- and third-tier suppliers won AI orders because the leaders were fully booked with NVIDIA work. If AI demand cools, the leaders get room to win back other customers, and overflow orders are the first to go.
TSMC shows what this looks like for a leader. Reuters reported that TSMC has told some China-linked customers to halt orders at 16nm and below unless they hold a U.S. export license. That capacity can be sold to other customers.
Prices are likely to move before volumes. Customers that placed extra orders to lock in capacity can cancel them once projects slip by a quarter, which weakens the case for this year’s price increases. For many suppliers, holding volumes while giving back price hikes would already be a good outcome.
I’ll watch whether the leaders’ monthly sales hold up while smaller suppliers slow. That would be the first sign the split has started.
Taiwanese CCL giant Elite Material (2383 TT) really put out a whole filing to say Nvidia has never been more than 20% of its revenue 😭🥶
The stock still closed 45% below its recent high on Nov. 17.
The post-Jensen era is so messy lol
ctee.com.tw/news/20261118700…
13
1,470
Google’s TPU v9 has an encouraging production update.
MediaTek told Morgan Stanley’s Asia Pacific Summit that it’s comfortable with how yields and output are progressing. It still sees EMIB-T substrate supply as key to the ramp.
MediaTek also said Google could use TPU v8t and v9 in rack systems for outside AI labs, in addition to using them internally. That could give both chips another source of demand.
Timing still matters. A later start would give v9 less time to ramp before v10 arrives, and the report didn’t give an updated shipment timeline.
Standard MLCC pricing is firming up, too. China’s Three-Circle (6951 HK) told Morgan Stanley’s Asia Pacific Summit that its average MLCC selling price rose 20% to 30% quarter over quarter in Q3 2026, mainly because it raised prices on mid- and low-capacitance parts.
As Murata and Samsung Electro-Mechanics shift capacity toward higher-end parts, buyers of standard MLCCs have started qualifying other suppliers. That’s why Morgan Stanley sees Three-Circle’s price increase as a positive signal for Yageo (2327 TT)’s standard MLCC business.
Three-Circle expects to end 2026 with capacity to make more than 70 billion MLCCs a month, up from about 50 billion at the start of the year. It still sees a shortage of high-capacitance parts through the first half of 2027 and says it has yet to close the technology gap in making those parts in smaller sizes.
Taiwan stocks fell 4.4% yesterday and were down more than 1,600 points early this morning. Japan’s Nikkei was down 4.1%, and South Korea’s Kospi was down 4.7%.
Taiwan stock market’s NT$10,000 club is now empty. As of this morning, only 30 stocks were still trading above NT$1,000 (around US$30.32) a share.
Let's not stare at our brokerage app today.
Luckily one of my favorite directors just dropped the trailer for his new film about ChatGPT and Sam Altman. What an interesting time to be telling this story and i'm excited to see his take on it.
Hope this gives y’all a little break from the markets like it did for me.
Our future is in their hands. Luca Guadagnino’s ARTIFICIAL stars Andrew Garfield, Yura Borisov, Monica Barbaro, Cooper Hoffman, Jason Schwartzman, Chris O’Dowd, with Ike Barinholtz, and Mark Rylance. In select theaters Christmas Day. Everywhere January 8.
Artificial - Official Trailer - In Select Theaters Christmas
2
29
1,799
I think Street forecasts for 2028 AI chip spending, aka the numbers behind $NVDA and $AVGO estimates, are too high for the capex that’s supposed to pay for them.
The market has GPU and ASIC spending growing about 46% in 2028, to roughly $1.0T. Consensus capex for the big five cloud providers and neoclouds grows only about 14%, to roughly $1.78T.
Put those side by side and chips rise from about 35% of that capex in 2026 to 44% in 2027 and 57% in 2028.
That implies spending on everything else, from racks and power gear to networking and buildings, falls about 11% in 2028.
I expect that spending to grow with the chips. New GPUs and ASICs need power, cooling, networking and floor space before they can earn anything.
To support the chip forecasts with chips holding 2027’s 44% share, 2028 capex would have to come in about 28% above consensus. I think it’s more likely to come in 5–15% below.
When the two numbers have to meet, I think the chip forecasts are the ones more likely to come down.
Some chips go to buyers outside these companies, so the ratio is approximate. A 13-point jump in a single year is still hard to explain that way.
The chip forecasts also leave less room for growth than the headline suggests. On the market’s own numbers, the 2028 increase in chip spending, about $318B, is already about 6% smaller than 2027’s $339B. At a 44% share, it would be about 70% smaller.
Fab and advanced-packaging expansion follows that yearly increase. I think a smaller increase shows up first in new equipment orders for 2028–29 capacity. 2027 tool shipments are still backed by orders already placed, with $LRCX guiding 2027 WFE above $200B.
I think 2028 AI capex is more likely to come in 5–15% below consensus.
OpenAI and Anthropic are growing revenue too slowly to cover the rent on everything being built for them. By my model, the two labs need about $15B of net new ARR every month through 2028 to pay their compute bills.
This year they’ve averaged about $10.5B.
I started from the landlord’s side. NVIDIA puts a gigawatt of Rubin capacity at about $60B.
Depreciate the IT gear over 5 years and the building over 14, charge a 9% cost of capital, add power and operations, and the owner needs roughly $14B a year in rent per GW to break even.
That’s on the low side. CoreWeave says its new 4–5 year contracts price at $20–25M per MW.
Consensus has the big five cloud providers plus neoclouds spending about $1.0T, $1.6T and $1.8T in 2026–28. After stripping out non-AI spend and capacity they use themselves, I estimate OpenAI and Anthropic will be paying about 70% of the rent on leased-out AI capacity by 2028.
Covering that would take about $441B of combined revenue in 2028. At this year’s pace, I estimate they reach about $347B, 21% short.
Monthly net new ARR is the number that matters, and it has been choppy even as the base grew.
Anthropic’s annualized revenue rose about ninefold this year, yet it added about $16.7B in April, around $3B a month in August and September, and about $10B in October.
Token prices roughly halve every year, so usage has to grow about 4x a year just to keep revenue on track.
GPU life is the one variable that closes the gap by itself.
CoreWeave said Ampere and Hopper pricing rose 30–50% quarter over quarter in Q3. If older chips keep renting at strong prices and servers last 7 years instead of 5, the revenue needed in 2028 falls to about $342B, just under this year’s pace.
That has only been tested while compute is scarce, so I keep 5 years as my base case.
Most 2027 spending is already under contract, so I expect the adjustment to land on 2028 expansion.
Oracle and the neoclouds look most exposed. Analysts estimate OpenAI accounts for roughly half of Oracle’s backlog, and by my math OpenAI alone needs about $90B of outside money to cover 2028 compute, while Anthropic has about $19B to spare. Neoclouds also borrow at higher rates.
Financing drives the size of the cut. My base case assumes OpenAI raises about $27B more. With no new money, the 2028 cut widens to 16–26%. With about $65B, consensus holds.
A 5–15% cut would slow 2028 GPU and ASIC spending growth from the roughly 46% the market expects to 24–39%.
Chipmakers would need less new equipment for 2028–29 capacity, so tool orders could soften first, as early as 2027.
Even after that cut, 2028 capex would still run about 50–70% above 2026. I still see AI demand expanding over the long run. Capex expectations got ahead of revenue.
Anthropic’s Q4 roadshow guidance implies the two labs’ combined ARR barely grows from October to year-end, which lifts the monthly pace needed in 2027 to about $16.5B.
Two straight quarters at that level would ease the pressure on 2028 capex and change my view.
$NVDA $CRWV $ORCL #OpenAI #Anthropic
Micron is talking about “3D DRAM” too, and I think it looks much closer to d-Matrix’s Raptor than to memory cells built vertically.
At the TomorrowX Summit, Micron described stacking DRAM directly on top of the GPU or ASIC compute die at the packaging stage. The 3D part comes from stacking dies, and the transistor structure inside the DRAM stays as it is. What you get is very high bandwidth and much less power spent moving data.
Two differences from Raptor stood out to me:
Micron still needs a custom base die to handle data transfer between the compute die and the DRAM. So I’m not sure yet whether the DRAM dies stay a standard design or become custom too.
The DRAM sits on top of the GPU/ASIC, where Raptor puts it underneath. That makes heat the bottleneck. Micron says it has to co-optimize with customers on compute hotspots, the DRAM stack structure, bonding materials and the cooling system.
Micron expects this kind of 3D DRAM to start reaching the market in 2028 to 2029 and become common around 2030 to 2031.
What I find most interesting is the business model. A custom base die plus thermal design done jointly with the customer means that once a customer adopts it, I think they are effectively tied to Micron. Switching costs go up, and so does Micron’s pricing power.
That is also the kind of business TSMC prefers.
If TSMC ever wants to make 3D DRAM, I think a model like this would suit it well.
A bit late to the party, but still want to share some of my thoughts on the rumor that TSMC may enter 3D DRAM at NVIDIA’s request:
(1)
What does “3D DRAM” mean here? The rumor leaves that unclear.
If it means a 3D structure built during front-end fabrication to increase storage capacity per unit of die area, I would see NVIDIA’s motive as backing another DRAM supplier to expand wafer supply and push down today’s high prices.
That seems a weaker fit for TSMC’s customer-specific foundry model, so I find this scenario less likely.
(2)
The rumor said the 3D DRAM project would be led by TSMC’s VP of advanced packaging technology development, which means that 3D DRAM could also mean vertical stacking during packaging, just like HBM.
NVIDIA’s HBM base dies are already moving toward custom designs manufactured by TSMC. The three major DRAM makers still produce the DRAM dies and handle stacking and packaging.
Asking TSMC to take over those two steps would basically mean asking it to manufacture HBM in direct competition with the DRAM makers. Again, I would see the motive as expanding supply to lower prices. This also seems a weaker fit for TSMC’s model, so to me the claim doesn’t work as well.
(3)
Another possibility is the kind of 3D DRAM in d-matrix’s “Raptor” inference chip, presented at this year’s Hot Chips.
Like HBM, it uses vertical stacking during packaging. But while current HBM pairs a custom base die with DRAM dies that retain a standard design, Raptor’s DRAM dies themselves are highly customized, removing the need for a separate base die as the data interface.
Both the DRAM design and stacking architecture are highly customized, making this inherently a made-to-order business. The chip designer leads the DRAM design, just like d-matrix does for Raptor, while TSMC provides fabrication and stacking services.
Competition with DRAM makers would be more indirect, and the impact on them would not necessarily be negative.
If NVIDIA is asking TSMC to manufacture this kind of custom 3D DRAM, that fits TSMC’s business model much better. Personally, I find this scenario more plausible.
(4)
Whatever “3D DRAM” means, TSMC’s business model is significantly different from that of DRAM makers.
If TSMC do enter the market, I would still expect it to manufacture against customer orders. It would not bear the same inventory sell-through risk as DRAM makers, so I would expect much less dramatic profit swings.
We just hit 5,000 followers!
This made our day. Thanks for reading and sharing what you’re seeing, too.
We’ll keep following the chip supply chain from here in Taiwan and sharing what we’re seeing on the ground. We’ll also introduce you to more of the Taiwanese companies that make it all work and show you where they fit in.
So great to have you all.
3
82
2,701
ASE Technology ($ASX) broke ground today on a NT$248.9 billion facility to expand testing capacity for high-performance chips. The new site is scheduled to be completed in phases in 2028 and 2029, Taiwan’s Economic Daily News reports.
ASE already has another testing facility under construction in the same industrial park. Its first phase is scheduled to start operating in April 2027, giving ASE an earlier window to add testing capacity while the larger site is still being built.
Once the new site has its test equipment installed and customer chips are qualified on those lines, ASE can test more chips in parallel and handle higher volumes as customers ramp production. Those are the milestones that will tell us when today’s groundbreaking starts to translate into capacity customers can use.
money.udn.com/money/story/56…
Coherent buys optical filters from Apogee Optocom (6426 TT) for assembly at its own plants, while using most of its in-house coating capacity for lower-end filters, according to Apogee’s Nov. 17 earnings call.
For links between data centers, the hard part is getting hundreds of coating layers right. Apogee says its 1.6T filters use about 150 layers, while some long-haul filters for links between data centers use 250 or more. Small errors build up with each layer, making it harder to produce filters that meet spec.
Coherent reported more than $200 million in quarterly data center interconnect revenue in fiscal Q1 and expects that figure to roughly double by the end of 2027.
On the supply side, Apogee says its 23 existing coating lines are already running at full capacity. It expects ten more lines to arrive over the first three quarters of 2027.
$COHR
Elon Musk said GPU power swings in AI training can break the shafts on gas turbines.
SpaceX pairs its AI data centers with Tesla Megapacks partly for that reason. Without power smoothing “at a wide range of frequencies,” he said, training runs are very difficult.
We see power smoothing becoming a fixed cost of every large training cluster, and Taiwan’s power-supply makers are built into the layers closest to the GPUs.
NVIDIA says a rack can swing from about 30% to 100% utilization and back in milliseconds, while power plants can take one minute to 90 minutes to respond to a sudden ramp. It also warns that repeated swings can cause resonance and stress equipment, the same problem Musk described.
No single device covers that range, so storage gets layered by speed. Capacitors inside the rack’s power supplies absorb millisecond spikes. Supercapacitor units cover sub-second to several-second swings. Battery backup units (BBUs) bridge about a minute, and site batteries like Megapack handle the slower swings at the turbines.
Taiwan’s suppliers sit in the rack layers.
• Lite-On (2301 TT) co-designed the GB300 NVL72 power shelf with NVIDIA, which built in electrolytic capacitors to absorb millisecond spikes. NVIDIA says the design cut peak grid draw by 30% versus GB200 in a Megatron LLM training test.
• Delta Electronics (2308 TT) covers three layers: AI rack power supplies, complete BBU systems, and standalone supercapacitor cabinets that sell for more than $200,000 each, according to our research.
• Dynapack (3211 TT) makes the battery packs inside Delta’s BBUs.
That rack-level content grows when Rubin-era 800V power racks arrive in 2027 and 2028 and BBUs move into them at 20 to 25kW per module. We estimate the BBU market inside IT racks and power racks rises from $3.3 billion in 2026 to $11.2 billion in 2028.
$SPCX $TSLA $NVDA
Silicon photonics wafer buyers have committed about $1.2 billion to reserve capacity at Soitec over the next several years, giving the company a firmer basis for expanding production.
Soitec received €86 million in cash deposits in the six months through September. The agreements also let Soitec see how much inventory customers are holding, which management says helps reduce the risk of double ordering.
Soitec now expects more than $300 million in photonics wafer sales in FY27, over three times last year’s level.
Management expects most of those sales to come from pluggable optical transceivers.
Soitec is already increasing output at existing fabs. It plans to equip its Singapore expansion in stages, with the pace and scale tied to how much additional capacity customers commit to.
Two U.S. ASIC customers could need up to 2 million sheets of CCL a month as they ramp purchases from Q1 2027. The initial demand is expected to come from boards for AI servers and network switches.
That could push M4 and higher grades above half of Nan Ya Plastics’ (1303 TT) CCL revenue, up from 30% in Q3 2026. The share is expected to approach 40% in Q4, ahead of the ASIC ramp.
Nan Ya’s CCL lines are running at over 90% utilization. Total capacity is expected to reach 6.48 million sheets a month by year-end and stay at that level through 2028. That puts the focus on what Nan Ya makes with those lines. Higher-grade laminates could account for a much larger share of sales even with overall capacity flat.
I think 2028 AI capex is more likely to come in 5–15% below consensus.
OpenAI and Anthropic are growing revenue too slowly to cover the rent on everything being built for them. By my model, the two labs need about $15B of net new ARR every month through 2028 to pay their compute bills.
This year they’ve averaged about $10.5B.
I started from the landlord’s side. NVIDIA puts a gigawatt of Rubin capacity at about $60B.
Depreciate the IT gear over 5 years and the building over 14, charge a 9% cost of capital, add power and operations, and the owner needs roughly $14B a year in rent per GW to break even.
That’s on the low side. CoreWeave says its new 4–5 year contracts price at $20–25M per MW.
Consensus has the big five cloud providers plus neoclouds spending about $1.0T, $1.6T and $1.8T in 2026–28. After stripping out non-AI spend and capacity they use themselves, I estimate OpenAI and Anthropic will be paying about 70% of the rent on leased-out AI capacity by 2028.
Covering that would take about $441B of combined revenue in 2028. At this year’s pace, I estimate they reach about $347B, 21% short.
Monthly net new ARR is the number that matters, and it has been choppy even as the base grew.
Anthropic’s annualized revenue rose about ninefold this year, yet it added about $16.7B in April, around $3B a month in August and September, and about $10B in October.
Token prices roughly halve every year, so usage has to grow about 4x a year just to keep revenue on track.
GPU life is the one variable that closes the gap by itself.
CoreWeave said Ampere and Hopper pricing rose 30–50% quarter over quarter in Q3. If older chips keep renting at strong prices and servers last 7 years instead of 5, the revenue needed in 2028 falls to about $342B, just under this year’s pace.
That has only been tested while compute is scarce, so I keep 5 years as my base case.
Most 2027 spending is already under contract, so I expect the adjustment to land on 2028 expansion.
Oracle and the neoclouds look most exposed. Analysts estimate OpenAI accounts for roughly half of Oracle’s backlog, and by my math OpenAI alone needs about $90B of outside money to cover 2028 compute, while Anthropic has about $19B to spare. Neoclouds also borrow at higher rates.
Financing drives the size of the cut. My base case assumes OpenAI raises about $27B more. With no new money, the 2028 cut widens to 16–26%. With about $65B, consensus holds.
A 5–15% cut would slow 2028 GPU and ASIC spending growth from the roughly 46% the market expects to 24–39%.
Chipmakers would need less new equipment for 2028–29 capacity, so tool orders could soften first, as early as 2027.
Even after that cut, 2028 capex would still run about 50–70% above 2026. I still see AI demand expanding over the long run. Capex expectations got ahead of revenue.
Anthropic’s Q4 roadshow guidance implies the two labs’ combined ARR barely grows from October to year-end, which lifts the monthly pace needed in 2027 to about $16.5B.
Two straight quarters at that level would ease the pressure on 2028 capex and change my view.
$NVDA $CRWV $ORCL #OpenAI #Anthropic
Starlink V3 could bring Taiwanese PCB maker Compeq (2313 TT) at least twice as much board value per satellite as V2 Mini, according to analyst estimates cited by Taiwan’s Financial News. That gives Compeq room to grow its satellite PCB business even if SpaceX deploys the same number of satellites.
SpaceX designed V3 for 1 Tbps of downlink capacity. Its upgraded antennas use a new generation of beamforming chips to handle more beams. More board layers give designers room to route signals, while low-loss materials help preserve signal quality at higher speeds.
Analysts expect the relevant boards to go from 18 layers or fewer on V2 Mini to 20 or more on V3, with an upgrade to M8 laminate. For Compeq, the share of V3 in SpaceX’s deployments will matter alongside the total satellite count.
fnc.ebc.net.tw/fncnews/stock…







