It’s pretty wild to see $NVDA crash -25% and lose close to $2T in valuation
After projecting 70% supply-constrained growth FY2028.
Also probably one of the worst times for $ANT to IPO at $1.4T ahead of Thursday…
Are there still people out there interested at that price? https://t.co/KYrQL2hyc4
New commentary today from SoftBank on Elon Musks post 4 days ago:
“SoftBank is defaulting on their turbine orders”
Analyst Question: “Any comment on that?”
SoftBank CFO Yoshimitsu Goto: “this is simply not correct at all. There is no such default”
This also mirrors what $GE CEO Scott Strazik stated about seeing no procurement cancellations or counterparty financing difficulties.
Pretty damaging allegations about the company financing $SPCX competitor (OpenAI + Stargate).
I wonder if there’s any more nuance these claims or if it’s false.
If you look at CW laser ASPs for AI data centers, the pricing differences are pretty substantial:
70mW: ~$3–4
100mW: ~$4–6 (Source Photonics channel checks).
With some premiums apparently going up to $7-$9 range with the recent shortage for pluggables.
120–150mW NPO: ~$15–20
350mW UHP CW: ~$40 (Nomura)
400mW UHP CW for NPO/CPO: ~$50–75 (Mizuho)
With former communication laser chips going from 20-50 cents.
The pricing and gross margins of these players go up with each architectural shift (60%+ for CPO lasers).
And go up even more depending how far down the stack you go (selling ELSFP was ~2.5x revenue content)
In terms of how I see the landscape now:
- $COHR is probably the best risk adjusted laser player now, for people who don't like uncertainty. Their new $20B LTA is very substantial.
- $LITE pretty negative now given the valuation overhang from $MRVL acquisition. Operationally very opsitive.
- $AAOI has extreme demand visibility (likely $6-$9B in new LTAs from just 1 quarter), but credible execution uncertainty.
They've also been focusing on pluggables + bottlenecked by DSPs, rather focusing on UHP CPO laser merchant markets.
But I think if markets see a new $5B-$7B LTA that they're in late stage discussion for... and see the $4.4B MC. Pretty sure that's a catalyst for recovery.
The demand doesn't disappear but timeline shifts back a few months.
AOI also signaled DSP as the main bottleneck, last quarter it was TIAs, so other different component players that are interesting.
- $SIVE has extreme capacity expansion (300m laser capacity targeted in total).
The economic implications that they won't disclose are likely enormous relative to their MC, even if they sell $3 70mW lasers to China pluggable players.
But there's a qualification uncertainty to fill the gap,
that markets have punished the company for. Especially when management communication amplifies uncertainty, despite the ongoing shortage.
Right now entire AI trade is going up and down together though.
我提出一个新颖的观点:
拿今天的市场和2000年科技股泡沫的比较其实是很可笑的
因为当时泡沫的重点是PE太高,高达100-200
而现在,PE最大也就20-30,
所以很显然,本次股价上涨的最高点,不可能达到2000年的巅峰时期
而是远在他之前就已经见顶
然后一路回撤了
也许根本就不会有望众瞩目的那最后一段blow off top
或者,blow off top已经发生了
再给大盘一年时间,就能看出真伪
谢谢大家
Missed the neocloud topic,
I think many neoclouds will be fucked up next year due to the tight financial conditions and will be acquired by CSPs with the asset price.
So that I’m not expecting a big disaster from the credit market.
25/25
I’m still very bullish on AI. Talk to people building it and you’ll find plenty of optimism. The bearish one might just be the guy trading the stocks 😂
NVDA the company and NVDA the stock are different questions.
Don’t let the narrative make you buy the top and only turn bearish after the selloff.
24/25
Hong Kong: I was long Xiaomi as a short-term trade.
Weaker USD and the narrative of China building DCs more easily could help. But BABA has a separate risk if tighter NVDA restrictions affect chip access.
I’m only trading around HK selectively. It still feels like one of the hardest markets to make money in 🥹
23/25
INTC was also on my original dip-buy list because I saw potential in advanced process and packaging.
Then earnings disappointed, the political support story weakened, and we discussed EMIB yield concerns. The exact yield talk was hearsay.
I think it needs time to prove itself. For short-term trades, I’d rather follow strength than wait for the turnaround.
22/25
That was why I originally liked LITE and the CIEN-type opportunity.
But the Marvell deal setup changed LITE’s price action. Merger-arb flows can tie the stocks together, like the CoreWeave/CORZ example we discussed.
I took LITE off my list unless the deal structure changes, such as shareholders rejecting it.
21/25
The optics thesis is more specific than just “AI needs more bandwidth.”
Getting one giant GW-scale power contract is hard. You may instead get smaller blocks of power across multiple locations.
That pushes distributed DCs, like the Google approach we discussed, and makes the latency between those sites important, especially for training.
That’s scale-across.
20/25
BE was one of my original buy-the-dip candidates.
Even before the election, DCs were being pushed to bring their own power instead of relying entirely on the grid.
I saw BE as one of the most credible solutions for that problem. The relative strength in the bounce reinforced why I was watching it.
19/25
For PLTR, what caught my attention was the Qwen + proprietary data fine-tuning example I discussed.
My takeaway was that an application company can use its own data and workflows to get better task-specific results at lower model cost.
Customers, data and distribution can matter a lot even as the frontier models improve.
The margin part is important.
With SaaS names like SNOW, the concern is AI features growing revenue while compute/model costs eat the benefit.
If PLTR can reduce dependence on expensive models, keep product quality and charge for the features, more of the value stays with PLTR.
That’s the economic moat I find interesting.
18/25
For TSLA, the first thing was robotaxi.
Based on the numbers I’d been looking at, I cited around $50M annualized revenue in Austin alone. That was the run-rate figure I was using, not realized full-year revenue.
If a city works and the next permits come through, the replication potential gets interesting.
The second TSLA point was robotics.
Recent model progress makes end-to-end hardware/software capability more important to me. If you rely on someone else’s hardware and only add a software layer, better general models can pressure your moat.
That’s why I like TSLA’s integrated position. Robotaxi and humanoids are the two application paths I’m watching.
17/25
SpaceX interests me because ground-based DCs face power and community resistance, including outside the US.
Could space DCs eventually become a more scalable route? I think it’s worth considering, alongside the communications infrastructure and model-related possibilities.
It’s a long-term possibility, and a personal preference. I wasn’t arguing it was cheap.
16/25
For the rebound, I generally prefer non-semi big tech. If trading AI hardware, I want relative strength.
In the bounce we discussed, AMD/COHR/BE looked strong. AAOI, which I’d been shorting, still looked weak.
Watch what the market keeps buying. A big drawdown alone doesn’t make something the best bounce candidate.
15/25
We disagreed on CLARITY.
I was very bearish on its prospects. Mr. Z wasn’t: he argued Democrats controlling both chambers could be less gridlocked than a split Congress, and they also need policy wins.
That was a real disagreement in the interview, not a shared conclusion.
Mr. Z put more weight on TGA liquidity and AI absorbing capital in private markets.
His question was how much liquidity Treasury would release, rather than treating crypto as only a political trade.
I agreed liquidity matters for both BTC and gold. My extra concern was the policy exposure.
11/25
Vol-control funds are another piece.
They size exposure using realized vol. The recent selloff raises that measure across the 20/60-day windows we discussed, so they can still be sellers even after price starts bouncing.
The market needs time to repair its positioning and internal liquidity.
10/25
Iran may feel like background noise to a stock investor. A vol seller has to care about the tail risk.
My view was that any escalation before Jan was especially relevant, before the new Congress adds constraints.
That uncertainty can keep vol sellers from coming back aggressively.
I read the market as still being in negative gamma. With vol sellers reluctant to return, moves could get amplified both ways.
The bounce didn’t have huge volume. My read was that put closing and vol crush were creating mechanical buying.
Price can rip before active funds regain confidence.
9/25
Year-end funds could be a headwind.
After seeing margin calls and funds shutting down, managers are more worried about protecting P&L than adding risk.
A Santa rally isn’t automatic just because they need performance. Some may use the rebound to reduce positions and close their books.
8/25
I feel we’re not far from a bottom, but I wouldn’t assume a straight V.
Compared with July, the funds I spoke with were much more bearish on fundamentals this time.
My preferred path is a bounce toward the 200D MA/old support, then a retest and hopefully a higher low or W.
Thursday’s low may hold. Another possibility is a test of the earlier pre-Iran high.
I’m watching the retest, QQQ reclaiming the 200D MA, and whether we can take back 680–700 before extending higher.
These are possible paths. Predicting the exact shape is basically fortune-telling 😂
7/25
From the interview through the Jan transition, I lean long.
The semi forward PE I was looking at had fallen to a little over 13x, around the 2022/low-teens comparison I mentioned.
After Thursday’s capitulation, semi and Mag 7 positioning was much lighter too.
I think chasing shorts there has poor RR.
What could take semis down another leg?
Late Jan/early Feb CSP earnings and 2027 CapEx guidance.
Money, power and the political constraints could produce guidance that semis won’t like.
If that bearish thesis gets confirmed, I could see another derating, potentially toward 10x. That was a scenario, not a price floor.
6/25
Anthropic: I’d heard from friends that it was willing to accept lower pricing to get the IPO done.
If it really comes at $1.1–1.2T, I’d buy without much hesitation. My own fair-value range was $1.6–1.8T.
The roadshow talk was secondhand. The buy condition is the actual valuation.
5/25
For NVDA, the tradable issue for me is tighter China export controls, including third-party routes through Southeast Asia and the Middle East.
My estimate in the interview was 20–30% exposure to Chinese demand including indirect channels.
That’s my estimate, not NVDA’s disclosed geographic revenue split.
I don’t think a full block of those channels was fully priced in.
My expectation was more scrutiny and tighter controls after the new Congress takes office in Jan.
So even after the selloff, NVDA’s earnings assumptions could still change. A low PE only helps if the E holds up.
4/25
On the midterms, I had bet on a Democratic sweep on Kalshi when the odds were around 60%.
But I got something else wrong.
I thought both parties needed AI, and local resistance to DCs was already bipartisan, so the election wouldn’t change the buildout that much.
I underestimated the governors.
The governor flips matter because in many states, the governor appoints the people overseeing energy regulation.
Power approvals can directly delay a DC. This can matter more immediately than legislation in Congress.
Both parties can support AI while actual projects get slowed by power, water, noise and local politics.
3/25
I cut basically all my semi longs first by the end of September. CPU, memory and optics were already very crowded.
Then I started building semi/Nasdaq puts.
The $70B OpenAI ARR story didn’t make sense to me. A jump from $40B+ to $70B needed explaining. The later number I was following was around $50B.
That expectation gap was part of the short thesis.
Then CoreWeave CDS widened and I started shorting neoclouds too.
I mostly ended that directional short trade around the middle of last week. What remained was mainly hedging.
The sequence matters: cut longs first, build shorts as catalysts developed, then cover as the setup changed.
2/25
Back on Oct 7, my framework was high rates + terrible breadth + crowded AI trades.
We were waiting for US10Y/MOVE to peak and roll over.
Breadth could repair through ITB/KRE/IWM/XLF catching up, or the AI leaders falling.
Over the past month, both happened.
1/25
As promised, English summary of my interview with @168X_Fortune and @168MrZ.
My stance at the interview: mostly covered directional shorts, looking for a rebound into Jan, still bearish on 2027 semis, very bullish on AI longer term.
A lot depends on the time frame and price.
closed NVDA/SMCI/AAOI short
And longing NVDA at 166 as it’s almost March low.
Will be out if losing 165 or rejected by 181.xx. Also will re-short NVDA if it gets rejected by 181.xx, final target 153.
I think it’s fair if we wipe out all China related revenue.
Came across Gabe’s awesome video and it well explained why the QQQ 680-700 zone I mentioned last week is so critical(looks like just consensus from ICT trading guys😂 )
Highly recommended.
@GAndersonTrades Wow great just came across this great video as my friend @acemoney21 recommended it.
Very clear and well explained. Thank you so much.
I’m holding the exactly same view as yours(I guess just because we are both doing ICT trading 😂)
DDR5 32GB RDIMM spot prices are pulling back.
9/28 $2,163.50
→ 10/5 $2,176.00(+0.58%)
→ 10/12 $2,185.00(Peak)
→ 10/19 $2,165(−0.92%)
→ 10/26 $2,165(0.00%)
→ 11/2 $2,025(−6.47%)
We’re seeing softer pricing in this segment of the server memory spot market. Server DRAM contract prices were still rising in TrendForce’s November 4 update.
$ANT’s headline IPO valuation is roughly $1.02T at $78/share.
If ARR hits $150B next year, here’s the math:
$150B ARR × assumed 20% net margin = $30B annualized earnings
$1.02T ÷ $30B = ~34x run-rate earnings, before further dilution.
Cheap or expensive? Your call.
**TSMC’s October revenue jumps 59.7% year over year**
$TSM (https://t.co/rPM4b3sBx5) reported October revenue of NT$586.89 billion, up 14.7% from September and 59.7% from a year earlier.
Revenue for the first ten months of 2026 reached NT$4.49 trillion, up 43.3% year over year. The October result gives TSMC a strong start to the fourth quarter.
TSMC official announcement https://t.co/t3uYi3LwMi
Keysight and Tower bring electrical–optical co-design into one workflow.
$KEYS has certified ADS Photonic Designer for $TSEM’s PH18 silicon photonics PDK, now available. Engineers can combine electronic circuits with physics-based photonic models for electrical–optical–electrical simulation, schematic design and layout inside ADS.
For 1.6T links and CPO, the value is evaluating how drivers, photonic transmitters, optical channels and receivers interact before fabrication. This could expose system-level performance problems earlier and reduce costly design iterations. Foundry-level verification and tapeout still use Tower tools.
Keysight announcement:
https://t.co/9B9QeGi7vR
SpaceX is building a 5GW AI data center in orbit by 2028.
72 Rubin GPUs per Kessel rack. 400kW. Liquid-cooled to space radiators.
Earth’s energy wall is about to get bypassed. Still sitting on the sidelines while @sssjeffpu sounds the charge?
For investors in Taiwan, the entire local market is essentially one giant AI trade. True diversification is a luxury that simply doesn’t exist here. When AI booms, the market rips; when AI tanks, everything sinks with it.
Nobody bothers digging into non-AI stocks—it’s pointless. Even names with zero exposure get dragged down whenever the broader market takes a hit.
That leaves a surprisingly simple playbook: a binary choice between being fully invested and sitting on cash. When risk management calls, cash is the only real friend.
Since de-risking my Taiwan portfolio on October 13, Liwei is back to buying aggressively, still heavily focused on testing and CPU supply chain plays—the two areas where Liwei has the highest conviction…hope everyone survived over this slaughter..
Similarly, Fable 5.2 uses 2.3x more output tokens than Opus 5.5 to achieve a marginally worse partial score on GDPx finance tasks. (2/3) https://t.co/VyPDDmgH4C
Even for domains where Fable 5.2 shines, it's still meaningfully more expensive. Although Fable 5.2 nearly triples Opus 5.5's score on GDPx software engineering tasks, the average cost per task also increases by roughly the same multiple. (3/3) https://t.co/mGjFX5D8xs
Fable 5.2 is very token inefficient for certain tasks due to its increased persistence and proclivity for spawning subagents. For example, on GDPx materials science tasks, Fable 5.2 costs 89% more than Astra and uses 4.3x more output tokens to achieve effectively the same score. The price difference vs Gemini 4 Carbon is even more startling at 7.5x. (1/3)🧵
HBM4 already doubled the number of pins from 1024 to 2048. The GPU's shoreline has limited additional space, so HBM5 keeps the same number of pins and makes each one faster to improve bandwidth. (2/6) https://t.co/mcGMmqYqVS
Step 3: The HBM PHY which drives the pins must also run twice as fast to get the data out. In HBM5 it moves to the edge of the base die, cutting the path to the GPU from ~5.5mm to ~2mm per Samsung. Same pins, ~2x the bandwidth. (5/6) https://t.co/Nn6qJQRDhH
Step 1: DRAM banks cannot work 2x faster, so HBM5 adds more of them. We estimate 2x the banks per DRAM layer to get there. (3/6) https://t.co/jzxorBRI1c
🚨 HBM5 Doubles Bandwidth Without Adding a Single Pin 🚨
HBM5 keeps the same 2048 pins as HBM4E and runs each one nearly 2× faster. Samsung's Hot Chips 2026 presentation shows pin speed going from 14 Gbps on HBM4E to over 28 Gbps on HBM5.
Faster pins are only half the story. To keep each pin fed, everything behind the pins doubles too: DRAM banks, TSVs, and an upgraded PHY on the base die.
Here's how it works, explained as a grocery store 🛒 (1/6)🧵
The catch: HBM5's PHY can run at more than 4x the power of HBM4E. To draw heat off the hot spot, suppliers are introducing a thermal "chimney" placed on top of the PHY.
We cover HBM bandwidth, cost and bit supply in our Memory Model and in "All About That Base: Custom HBM and the New Memory Economics" Subscribe to learn more!
https://t.co/jGrJoiNm7Z (6/6)
Step 2: More data requires more lanes to move it, so the TSVs, which carry data and power up and down the stack, double as well. (4/6) https://t.co/L5w4laZvLK
However, this increased persistence comes at a cost. Compared to Fable 5.1, the average cost per task for Fable 5.2 is 31% higher ($67.16 vs $88.15). This is despite Fable 5.2 being 20% cheaper per Mtok. (6/7) https://t.co/sttqVAbCxW
Interestingly, Fable 5.2 scores about the same as Sonnet 5.5 on RSI. Anthropic explicitly limited advancements in AI research during training, and it's cool to see GDPx accurately capture both the SWE improvements and plateau in RSI abilities. (4/7) https://t.co/nD8dX3Q4Fz
Overall, we think it's clear why Anthropic named this model Fable 5.2 and not Fable 5.5. It's a marginal improvement, incomparable to the original Mythos release, and likely far behind what Anthropic has internally. It will be interesting to see if they release anything more impressive before their IPO.
For more details on Fable 5.2's GDPx results, checkout our leaderboard👇️ (7/7) https://t.co/OBBaXzmC2Y
After analyzing the rollouts, there is one primary reason behind Fable 5.2's improved score: it's much more comprehensive when testing its work. Fable 5.2 will often generate hundreds or even thousands of test cases, whereas Opus 5.5/Fable 5.1 stopped at just ~20.
Relatedly, Fable 5.2 is also much more eager to launch subagents for independent re-derivation and auditing. (5/7)
Materials science (+5.3%) and bio (+7.7%) also saw meaningfully improvements, while others like quant and hardware design were more a less a wash. (3/7) https://t.co/TRHkdHh22u
The gains are primarily concentrated in chip design and software engineering. Fable 5.2 beats Opus 5.5 by 10.8% and 15.5% respectively. (2/7) https://t.co/j6MKkz051g
BREAKING NEWS🚨: Today, Google announces TPU 8tx for confidential inference and TPU 8i Max!
This builds on the rapid progress they are making across the TPUv7 inference stack, much of which can, at a high level, be transferred to the TPUv8 series, including Muse Spark Open, @sgl_project and @radixark.
TPUv7 MOGGING GB300 NVL72!