@MarkTheLaird Just in general for laser players, from Lightmatter's CEO, regarding scale up CPO:
8 optical engines/GPU introduces 100x more lasers than scale out switching. More conservative estimates were 10x laser content from AOI's CFO.
Along with ASPs for those lasers going higher (eg. 10x ASPs than pluggables for single emitter UHP lasers).
From what I've seen yearly capacity expansion for lasers generally increase around ~3x Y/Y like $AVGO projections, or ~5x over 2-3 years like Furukawa's.
This is what I mean by optical content like lasers + requirements (more InP substrates), rises per generation and bottlenecks on UHP ranges might persist through 2030 despite aggregate capex decreasing.
Which is what I'm trying to argue here.
Future CPO laser competitors like Furukawa/AOI/Shija are more nuanced, since AOI is focused on 800G/1.6T transceivers (which faced capacity expansion delays + DSP Bottlenecks), but they have capability for 400mW+ ranges + ongoing CPO qualifications.
I don't think transceivers have the same durability though in capex reduction scenarios, which is why I've been arguing the valuable optionality AOI has to focus on merchant lasers is mispriced.
A lot of it is execution/qualification rather than demand, $LITE + $COHR are the most de-risked with confirmed long term visibility.
But entire AI trade is moving up and down together.
Well, most of the time projections are noise and reactionary.
Like right now, sky-high projections for $ANT $700-$800B ARR over next few years is stirring up hype that I'd argue gets revised down later.
I do think it's important to model in capex forecasts dropping past 2028, due to the current reflexive nature of AI selloffs, rate hikes, etc.
The drawn out duration of Iran War doesn't help, Jensen leaving $NVDA certainty draws concerns over Nvidia's ecosystem.
But any catalyst, like if Iran War ends, can also change sentiment in a heartbeat (since bearish positioning is overwhelming).
- It's also clear financing has been tightening, since OpenAI only managed to raise $11B of their original $30B target so far.
- Markets are reacting enthusiastically to companies like $TE doing a $100M ATM today (nuking valuations)
- Although the "guppy" sector is probably the most sensitive to financing + rate changes.
On bright side... Neoclouds like $NBIS or $IREN seem to be adjusting by getting 100% capex prepayments or 90% GPU prepayments from their counterparties.
What I'd argue though is lower aggregate capex does not apply uniformly in supply chains.
- Mid-high range lasers for NPO/CPO are still sold out, as seen with $LITE now starting negotiations for 2030 capacity.
- MLCCs are seeing further tightening as content usage goes up for AI servers
- ABF substrates to InP substrates not really going anywhere.
If content per system goes up 10X-15X, capacity expands at 2-3X Y/Y, and capex drops 28%.
And content is so cheap that you can triple order and screw over competitors with wafer supply...
I'm pretty confident you'll see acceleration upstream in bottlenecks from optical components to substrates despite potential aggregate capex dropping.
Another rant is:
-I keep seeing this argument that "low P/E is a warning for cyclical that's going to crash" but these are just clown takes
Nobody seems to analyze duration scenarios like capacity expansion vs. content changes vs. demand. to actually see how long those earnings last.
If everyone seems to be exiting DDR2 lines, there's probably going be more tightness in 2027 despite consume demand weakening, even if it looks like it's low P/E.
- Also for concerns about CPO "delays", the same components like lasers are used for NPO. You'll still see 150-200mW NPO lasers or ELS NPO using the same 300–400+ mW lasers as CPO.
We'll just see acceleration with NPO as a bridge if CPO gets delayed and higher range lasers likely to remain in a severe shortage.
- Feels like markets have also completely discredited the merit of take/pay price floor LTAs as well that companies like SK Hynix to Samsung have.
If contractual agreements hold up, memory players look very undervalued.
TLDR:
- I think it's important to model in capex decreasing to numbers like -30% in hypothetical downside sceanrios.
Given AI profitability, it's clearly not sustainable for hyperscalers to aggregately increase capex $500B or $1T+ Y/Y, and that's never been a case I've argued.
- This is also not the same thing as demand destruction that markets are overreacting to now (which haven't happened yet).
- We can see demand revisions for GPUs, ASICs, and heavier cost parts of the supply chain, while acceleration increases elsewhere.
The nuances market miss by indiscriminately selling everything is going to present compelling opportunities when the dust settles. I think in the end institutions are probably going to end up owning more of the same names retail discovered early on.
Anyway, I think it's a good idea to try and survive until more earnings comes out.
I don't know what it takes to convince the Taco Bell board member or Washington at this point?
$NVDA is not the same thing as Apple or McDonalds.
You also cannot replace Jensen with ARM CEO or Palo Alto Networks CEO... then expect local Taiwanese factories to move as fast. https://t.co/kSoTRrIZln
FYI if you're going to short Unitree and humanoids regarding OpenAI news.
"Disbanding" isn't bearish on humanoids, just misinterpreted.
I'm hearing rumors that OpenAI is shifting more toward M&A over in-house. They did try buying 1X last year as a reminder. https://t.co/4PgdP4HiZB
@invest_create Not even joking I heard a hedge fund manager the other day say
“I don’t like $INTC” not a fan of memory.
Majority won’t understand the nuances
@ste11aaaa_ Right now they’re all being treated indiscriminately which is annoying.
I’d expect GPU/accelerator vendors to be hit hardest, assembly/ODMs/OSATs, etc.
Valuation cut in stuff like equipment/test.
Upstream components like medium-high range lasers for NPO/CPO, MLCCs, certain substrates, etc to see demand acceleration
As content per system increase a lot for future architectures, they’re cheap + relative to global capacity expansion.
Also not quite sure why Neoclouds are treated as “it’s over”.
Their growth + capex slows but they’ll probably be more profitable
If DC slowdowns happen, compute prices rise, and they monetize their existing capacity.
I think AI is one of the few industries where investors celebrate deteriorating supply conditions.
MLCC lead times are extending per new Trendforce data… now “very tight” at 40W.
Content per system keeps going up as well (eg. $AMD using +632% more of a part for MI450 boards).
Yet suppliers from Murata to Samsung Electro-Mechanics are getting sold off to oblivion.
I will continue to make the argument that we can continue to see demand acceleration in certain bottlenecks…
Despite potential overall capex reductions past 2027. Would not conflate the two like what markets are pricing in now.
@AsgardArru I just find it interesting that everyone likes to use the Warren Buffett quote: "Be greedy when others are fearful"
But when a crash actually happens, everyone makes fun of folks like Morgan Stanley when they say "Maximum Overweight"
Not saying I agree with this positioning, but I just found their stance interesting enough to share.
Morgan Stanley sends a note to clients using the term:
“Max Overweight” for global equities.
Arguing that AI capex cycle remains intact and sustains corporate earnings into 2027.
They told clients:
- hold less cash, trim commodities and exposure to bonds
- move Europe/Japanese equities to overweight
- Global stocks trade ~15x fwd earnings (MSCI ACWI), the cheapest since 2023
- AI theme is expected to spread beyond the hyperscalers
- Global earnings growth is forecast at 13% in 2027 and 8% in 2028
But flagged:
- Inflation could force more tightening than the two further Fed hikes
- AI capex cycle may not last as long as assumed
Earlier this week they argued:
“The hedge fund deleveraging unwind has been brutal since October.
While retail investors remain calm and long-term oriented, levered institutional hedge funds have acted as ‘chickens with their heads chopped off’ (IMO),
moving daily stock prices around at nonsensical speeds.”
Anyway the most recent time I found this term used was 14 years ago June 8th, 2012.
Where they called “maximum” overweight on emerging-market equities.
Usually I treat this as noise but TIL “max overweight” was a term used.
@Jonatha29815168 SA was super stubborn and overconfident on the power supply and capex growth months ago when I said that’s bullshit
Now it joins the FUD when everyone already sold semi😂
Please stop asking me why everyone is bearish now😂
NQ probably will down another 200 points from 27650 now and if hold & reversal at 27450 then that’s it 😓
Joined a happy hour event of @FundaAI tonight.
100% bears among the HF folks.
I shared my views that we shouldn’t be too bearish towards AI trades at the moment, and then none of them want to talk to me as they probably think I’m an idiot.
@AG3747373132582 How is this fear porn? Telling people to stay away from $NVDA’s supply chain while being bullish on ASICs like TPUs—isn’t that just rational analysis?
There’s talk that neocloud credit is nearing breaking point. Liwei ran the numbers.
Across five listed neocloud and AI-infrastructure companies, $CRWV, $NBIS, $IREN, $APLD and $CORZ, reported borrowings total approximately $81.6B, excluding leases and undrawn facilities.
If all five defaulted and creditors recovered 50%, the modeled loss would be $40.8B. Every 10-percentage-point decline in recovery adds another $8.2B to that scenario.
For context, the FDIC’s end-2025 estimate puts the cost of the 2023 U.S. bank failures at $34.7B. Credit Suisse’s historical AT1 write-down adds roughly $17.8B, bringing those selected components to about $52.5B.
The $40.8B stress scenario is roughly four-fifths of that partial benchmark. A broad neocloud default could therefore produce losses measured in tens of billions, although these figures do not establish its probability or timing.
The pressure points are cash interest coverage, refinancing deadlines and collateral recovery. With CoreWeave carrying over half the sample’s debt, its financing terms deserve particular attention.
Recovery assumptions also need a tougher test: how much would GPU and data-center assets fetch if several operators needed to sell simultaneously? A financing squeeze could force those sales and depress recoveries, increasing losses precisely when lenders expect collateral to protect them.
@imnotharsh @dylan522p This is very concerning…
1. Pricing for the Feynman generation was lowered even further beyond the HBM cost adjustments, as SemiAnalysis believes NVIDIA will have to accept lower gross margins to stay competitive. Feynman dropped from roughly $110,000 to $78,000, Feynman Ultra from roughly $175,000 to $98,000, and Feynman Next from roughly $191,000 to $118,000.
2. The cuts extend beyond GPUs to include Groq LPUs and standalone CPUs as well.
SemiAnalysis’s sharp downward revision of global AI accelerator demand, shipments, and capex forecasts, with the biggest cuts starting in 2028 (e.g., 2028 AI capex lowered 28% to $758B).
• NVIDIA ($NVDA) bears the heaviest pressure: 2027–2030 shipments cut 22–66%, pricing on Feynman GPUs also reduced, and it faces highest exposure to smaller Neocloud customers plus potential export-control risks.
• Other players fare better—Broadcom ($AVGO) sees only minor ASIC revenue impact due to strong financing support, MediaTek faces larger TPU cuts, AMD experiences milder reductions thanks to hyperscaler ties, while Amazon Trainium remains largely unchanged in the near term.
Detailed analysis ⬇️⬇️
Post-Jensen Huang era is already showing cracks.
NVIDIA’s Vera Rubin volume ramp is reportedly delayed from end-2026 to H1 2027 due to HBM4 supply bottlenecks.
GB300 remains the main volume product for now.
Impact on Taiwan supply chain is mixed:
• Server ODMs such as Hon Hai (Foxconn, 2317) and Quanta (2382) are largely unaffected. They have already passed the learning curve and can deliver in volume.
• Newer components that were counting on riding the Vera Rubin ramp — including next-generation power solutions (800VDC), high-end substrate materials (M9), and optical communication components — face potential delays or weaker near-term demand.
Bigger Question: Will NVIDIA’s aggressive 1.5-year material & platform upgrade cycle (memory, substrates, optics) slow without Jensen?
Eyes on Feynman 2028 specs.
With Jensen’s exit from $NVDA, diversification just became non-negotiable.
OpenAI was running ~90% of its training on NVIDIA GPUs — now they’re scrambling to shift.
Google TPUs and MediaTek are the clear winners in this transition.
Taiwan’s $NVDA supply chain got destroyed today.
Yunnan Germanium’s sub Yunnan Xinyao signed a 5-year framework deal with $AAOI for 4-inch n-type InP substrates.
Projected volume: 180k wafers/year (2027–2031) → total ~900k.
2027 contract value: $194.4M (ex-tax) = 122% of Yunnan’s 2025 revenue.
InP is the key base material for EML lasers & high-speed PDs in 800G/1.6T optical modules — a critical bottleneck in the AI optical interconnect supply chain.
$COHR $LITE $MRVL
https://t.co/X59ltXBVzM
Per @Grok If Jensen comes back, priority clean-out list:
• Colette Kress (Interim CEO)
• Harvey C. Jones (new Board Chair)
• Independent directors who pushed the ouster
• Kimberly Powell (VP Healthcare, internal affair)
• Anyone on the special committee that handled the resignation
• Board members who stayed silent or backed the transition
Power structure first, then the internal optics problem.
https://t.co/h9sCw9A0Cu
ALERT 🚨: Trainium3 Neuron profiling & trace stack is so fine-grained and precise that it can generate pixel art of entire videos!
The video below is an actual profile running on Trainium3 DMA engines.
We will open-source the entire SKILLS.md later this month. https://t.co/Yopt92VTvt
idclip? noclip? IYKYK. The tear in teardown doesn’t always have to be destructive. Part of the Kirin 9050 Teardown. Reach out to learn more!
sales@semianalysis.com \| https://t.co/NVaBFeGVWv https://t.co/za6gDuuNX0
The devil’s advocate case for letting China rent American compute.
“If China has access to compute outside of China, it’s paying humongous margin dollars to the US ecosystem. Nvidia makes about $7 million of profit on every megawatt deployed. The cloud provider makes another $7 to 10 million at today’s pricing.”
“If ByteDance wants to compete with Meta on video recommendation globally, and you don’t let it buy the chips but you do let it rent them, Meta potentially has a fourth of the cost of compute.”
“Application-layer companies running on open models add up to well north of $10 billion in ARR. DeepSeek makes less than Baseten and Fireworks do serving its models. America is making so much money off Chinese models.”