TL;DR:
1. US stocks have lost roughly $6 trillion in market value over 45 days, as AI-related selling intensifies.
2. AI CapEx is hitting record highs, raising serious concerns about free cash flow and returns on investment in 2027–2028.
3. Corporate earnings remain strong despite the selloff. The market is pricing in a much weaker future than current fundamentals suggest.
4. Political uncertainty and potential AI regulations are adding pressure to tech valuations.
5. Higher interest rates are making AI infrastructure financing more expensive, although potential monetary easing in 2027 could change the picture.
6. NVDA and AMD face company-specific uncertainty, but Kobeissi argues that much of the downside may already be reflected in valuations.
7. Tech valuations have compressed dramatically. The real question is whether investors still believe future earnings estimates.
8. Extreme fear, hedge fund selling, and liquidity pressure surrounding Anthropic's IPO have amplified the market decline.
9. Kobeissi doesn't believe the AI bubble has burst. They see this as a correction, although further downside remains possible.
10. The biggest opportunity may be in quality tech companies with strong balance sheets, accelerating earnings, and durable competitive advantages. The market is pricing in plenty of bad news, but very little room for positive surprises.
The new BIS rules are shutting Chinese chipmakers out of $TSM advanced nodes.
But I think the market is missing something.
For $NVDA, this isn't necessarily about losing revenue. It's about losing potential upside.
With demand already exceeding supply, the real damage may be to future earnings upgrades, not today's sales.
1
48
HBM4 prices could jump 1.5-2.7x in 2027, and customers are already shifting from 12-high to 8-high stacks.
That's a pretty big signal.
AI demand isn't slowing, but memory costs are getting harder to absorb.
$MU gets the pricing power. $NVDA, $AMD, and $AVGO face a tougher margin equation.
The next AI bottleneck might be profitability, not compute.
3
104
TrendForce just moved GPU supply from "balanced" to "starting to tighten" as the Rubin transition kicks in.
-> MLCC lead times: 38 weeks
-> DRAM lead times: down to 19 weeks but still very tight, especially 32GB and 64GB RDIMMs
A tiny capacitor taking almost 9 months to arrive can hold up an entire AI server. $NVDA, $MU, Murata, Samsung, Electro-Mechanics, Yageo.
Interesting… new TrendForce lead times (normal lead time in brackets):
- DRAM: 18 weeks, down from 19 [8]
- ABF substrate: 48-56 weeks [12]
- HDD: 50 weeks
- MLCC: 38 weeks
- GPU: 30-45 weeks
Memory is loosening a tiny bit, but substrates are still 4x+ normal. The squeeze just moved downstream. Ibiden, Unimicron, SEMCO, Murata.
Macro Voss retweeted
$AAOI เซ็นสัญญา 5 ปีกับ Yunnan Germanium เพื่อซื้อแผ่น InP (วัตถุดิบทำ laser)
ตอนนี้ใครคุมวัตถุดิบ InP ได้ก่อนจะได้เปรียบมากๆ ในเวฟ optics รุ่นใหม่ๆ ในอนาคต
และดีลนี้ไม่ได้ดีแค่กับ $AAOI เพราะอีกเจ้าที่เป็นต้นน้ำ InP อย่าง $AXTI ก็อาจจะได้รับความสนใจด้วย
$AAOI signs $194.4M/year contract with Yuunan Germanium.
This is for 2027 alone, with a 5 year contract. Given an implied total of ~$970M in InP substrates over 5 years (before price hikes)
Hmm just some thoughts:
- AOI laser capacity expansion seems bigger than what’s disclosed based on procurement patterns?
180K wafers seems like an extreme amount compared to what’s needed for lasers within their own transceivers… if they can consume all of that
So maybe we’ll hear about a new UHP laser contracts in earnings, after $LITE announced a $5B contract.
- read through on $AXTI is positive, given they’ll probably be #1 in InP substrates manufacturing scale in the future + have contracts with $LITE and $COHR.
- positive AOI was able to secure capacity during the bottleneck
More spend is probably the boogeyman in the AI trade, but it’s a good signal for earnings.
Macro Voss retweeted
TrendForce just moved GPU supply from "balanced" to "starting to tighten" as the Rubin transition kicks in.
-> MLCC lead times: 38 weeks
-> DRAM lead times: down to 19 weeks but still very tight, especially 32GB and 64GB RDIMMs
A tiny capacitor taking almost 9 months to arrive can hold up an entire AI server. $NVDA, $MU, Murata, Samsung, Electro-Mechanics, Yageo.
TrendForce just moved GPU supply from "balanced" to "starting to tighten" as the Rubin transition kicks in.
-> MLCC lead times: 38 weeks
-> DRAM lead times: down to 19 weeks but still very tight, especially 32GB and 64GB RDIMMs
A tiny capacitor taking almost 9 months to arrive can hold up an entire AI server. $NVDA, $MU, Murata, Samsung, Electro-Mechanics, Yageo.
Macro Voss retweeted
The bottleneck isn't chips anymore. It's peak power.
OpenAI just launched "Grid Protectors", where future data center campuses cut their power use when the grid is stressed.
First site: Effingham, Georgia, with Georgia Power. The campus can shed up to 1,000MW on request, and OpenAI pays for all the power infrastructure itself.
Flexible load means faster grid hookups… which means more demand for on-site batteries, power gear, and backup generation. $SO, $VRT, $GEV
The bottleneck isn't chips anymore. It's peak power.
OpenAI just launched "Grid Protectors", where future data center campuses cut their power use when the grid is stressed.
First site: Effingham, Georgia, with Georgia Power. The campus can shed up to 1,000MW on request, and OpenAI pays for all the power infrastructure itself.
Flexible load means faster grid hookups… which means more demand for on-site batteries, power gear, and backup generation. $SO, $VRT, $GEV
$TSEM just printed a record quarter.
-> Q3 revenue: $526M, +33% YoY
-> Q4 guide: $600M
-> Silicon photonics alone targeting a ~$1.1B annual run rate in Q4… 5x 2025
-> Customers extended contracts into 2029 and prepaid another ~$305M in 9 months
When customers are prepaying just to reserve capacity, that's what a real chokepoint looks like. $GFS $COHR $LITE
Macro Voss retweeted
Holy sht, it's official! $MRVL and $LITE are combining.
All-stock deal, $115B enterprise value:
-> 4.5 $MRVL shares per $LITE share
-> ~$1,132/share, a 23.9% premium
-> They're pitching a $550B addressable market and $1B+ in annual cost savings
-> Expected to close H1 2028
Marvell is pulling the laser chokepoint (InP/EML) in-house. $AVGO now faces a full-stack rival… and every other optics buyer may need a second source. Makes $COHR more strategically valuable overnight.
Holy sht, it's official! $MRVL and $LITE are combining.
All-stock deal, $115B enterprise value:
-> 4.5 $MRVL shares per $LITE share
-> ~$1,132/share, a 23.9% premium
-> They're pitching a $550B addressable market and $1B+ in annual cost savings
-> Expected to close H1 2028
Marvell is pulling the laser chokepoint (InP/EML) in-house. $AVGO now faces a full-stack rival… and every other optics buyer may need a second source. Makes $COHR more strategically valuable overnight.
$SKHY just held its first ever Partners Day in Japan.
20+ Japanese suppliers showed up: Shin-Etsu, Sumco, JSR, Namics, Tokyo Electron, Screen, Disco, Advantest…
They shared the mid-to-long term roadmap so suppliers can plan capacity expansions.
With HBM sold out through 2029, the next chokepoint becomes Japanese materials, wafers, dicing and test tools.
Interesting… analysts estimate $INTC EMIB packaging yields are way behind CoWoS.
- EMIB overall: ~60-65%
- Silicon bridge embedding into ABF: ~25-30%
- $TSM CoWoS: ~98%
They also argue the 7.2% SOX drop was about Intel's guide and hyperscaler FCF worries… not AI demand.
So the CoWoS alternative isn't ready yet. $TSM still owns the packaging chokepoint, which fits the CoWoS price hike talk.
Oh look… Taiwan's biggest PCB makers are now personally flying to Japan just to get in line for equipment.
Zhen Ding, Compeq, and Unimicron execs went in person… and only cut tool lead times from 30 months to 28.
Low CTE glass cloth for high-end substrates is still short too, since new suppliers need customer qualification first.
New capacity is ~2.5 years away. Japanese suppliers like Nittobo hold the pricing power.
Woah… Korea's October chip exports just printed $56.44B. That's +259.4% YoY.
Second highest month ever, only behind September.
And it's not just memory:
- Computers (mostly server SSDs): +496.6%
- Power equipment (transformers, cables for AI DCs): record $1.99B, +74%
- DDR5 16Gb contract price: $40 in June -> $50 in Oct
Memory AND power still sold out. Bullish $MU, $SKHY, Samsung, LS Electric, HD Hyundai Electric ???
Bro. I think the next major AI bottleneck might not be GPUs.
It might be Capital!
Everyone is focused on GPU shortages, power constraints, and optical supply chains.
But there's another bottleneck I think markets are underestimating:
What happens when AI demand is still exploding, but the cost of financing new capacity becomes too expensive?
Companies like CRWV and NBIS could have billions in contracted demand. But having customers doesn't automatically mean you can finance the infrastructure to serve them.
Especially when you're building capital-intensive AI datacenters with debt.
What I'm watching:
CRWV / NBIS refinancing costs
AI datacenter project delays
Hyperscaler financing structures
GPU-backed SPV financing and debt terms
My thesis is that 2027 could be the year AI projects start getting delayed because of financing constraints, rather than GPU shortages.
The interesting part?
AI demand could remain incredibly strong while some infrastructure providers struggle to build enough capacity to meet it.
And that would make access to cheap capital another competitive moat.
Maybe the next AI infrastructure winners aren't just the companies with GPUs or power.
They're the ones that can actually afford to build.
$TSLA and $SPCX are in talks with $TSM regarding US-based production capacity.
The likely model involves a facility where $TSM retains the technology and holds a majority stake,
rather than a scenario where $TSM simply transfers the manufacturing know-how for $TSLA to handle production independently.
Actual production is expected to commence around 2030 or later.
In short: $TSLA and $SPCX will still rely on $TSM to manufacture their chips, rather than establishing a competing fabrication plant.



















