katz

Global Macro & Value Investing Analyst
Regular personal portfolio updates
Research coverage: Global Macro, AI Compute, Semis, Robotics, Space Industry
DYODD
NFA

[katz @katzspx]

https://t.co/R8yywT9g71

[katz @katzspx]

@jukan05 Very aggressive timeline. $YMTC gets billions from IPO for capacity build‑out, yet equipment and export restrictions remain the biggest bottleneck to dethroning Samsung and $SKHY in NAND.

[katz @katzspx]

From an economic‑interest perspective, the photovoltaic industry has long relied on government subsidies. Enterprises struggle to turn consistent profits while trapped in relentless internal price competition.
From Beijing’s policy standpoint however, China will launch a new round of capacity‑reduction initiatives during the 15th Five‑Year Plan. Low‑cost photovoltaic hardware will be deployed across western provinces to build clean‑energy infrastructure. Inexpensive panels also accelerate the penetration of advanced PV technology into ordinary households. I could not wrap my head around this logic at first, until I came across this picture.

[katz @katzspx]

@Yerocode0 Human‑mechanical ascension incoming.
Neuralink(consciousness upload)+Optimus(mechanical body)+Grok(Jarvis AI).

[katz @katzspx]

Great breakdown. What stands out to me is this paradigm‑shift: revenue growth is no longer only driven by transceiver unit volume, but by surging per‑unit content as we move toward 3.2T.
The InP supply bottleneck is the shared constraint across $SMTC, $AAOI and the whole optical ecosystem. Even with fab capacity ramping 3‑4x, it remains questionable whether expansion can keep pace with hyperscaler pull‑forward demand.
The margin profile looks exceptionally strong post‑divestiture, and healthy free‑cash‑flow reduces dilution risk. Still, one wildcard: can they keep passing on rising InP material costs if competition eventually intensifies?

[katz @katzspx]

@jimcramer Testing performance is meaningless in the face of mass production

[katz @katzspx]

This argument misses the broader trend. The AI accelerator market is bifurcating: hyperscalers build custom silicon for their own workloads, and $NVDA dominates the general-purpose market. There is no obvious middle ground where $CBRS can insert itself simply because OpenAI launched a chip.
If anything, Jalapeño validates the custom-silicon thesis, which makes hyperscalers even less likely to rely on external vendors like $CBRS. The competitive pressure intensifies the build-vs-buy decision toward “build.”

[Jukan @jukan05]

I came across an argument that Jalapeño might actually be bullish for Cerebras. The idea is that Jalapeño puts more pressure on Meta and AWS to respond to OpenAI, which could push them to adopt Cerebras more aggressively.

I’m not sure I buy that argument, though.

[katz @katzspx]

I think $GFS’s value lies in the silicon photonics ramp that only fully materializes in 2027–2028, not near-term trading catalysts. The $4 EPS for 2028 and long-term $6 earnings power are predicated on margin expansion from richer optical mix, but the market may not price this narrative until silicon photonics revenue meaningfully inflects. That’s why betting on it over a 6-week window is extremely tough, even if the long-term math checks out.

[katz @katzspx]

@Mr_Derivatives I don’t think these purchases by Jack Ma and other insiders will change much for $BABA. Insider buying is a signal of confidence, but it cannot offset the larger fundamental and macro overhangs weighing on the stock.

[katz @katzspx]

I think the real contradiction lies in the different time horizons analysts and investors use to evaluate assets.
Most investors deploy real capital, so emotional reactions are inevitable when facing sharp drawdowns over the short 1–6 month window. In contrast, sell-side analysts and institutions typically look out 12–24 months.
Let’s get one thing straight: we are living through an AI revolution, analogous to the Industrial Revolution and the Information Revolution that came before. All industrial restructuring carries lag. Widespread adoption only becomes possible once technological breakthroughs drive down costs. Right now, hardware costs keep rising alongside demand, which is actually unhealthy for the long-term development of the AI industry.
You need to pick an asset that can survive the AI revolution, wait out industrial restructuring and its inherent lag. That way you won’t lose conviction in AI equities.
All great investing goes against human nature！

[Nicholas Mugalli @RealNickMugalli]

If you’re a semi investor losing faith in the AI trade, bookmark this and read it line by line.

Raymond James just upgraded $AMD this morning to a Strong Buy with a massive $641 price target. This is the fundamental roadmap for where compute is heading.

They expect the server CPU market to explode to $201 billion by 2030, driven by an insane 44% CAGR as hyperscalers scale up for agentic AI workloads. AMD sits in the sweet spot here, offering the cleanest mix of earnings leverage, pure play data center exposure, and market share gains against legacy compute.

Nothing in the operational thesis has changed. Hardware demand is accelerating, and the noise right now is just temporary macro sentiment…stay locked in!

[katz @katzspx]

My take:
As hyperscalers take on debt more and more frequently, their debt costs keep climbing. The sharp jump in $ORCL and $AVGO CDS perfectly illustrates the hidden leverage risk under the AI capex boom. Demand for chips stays strong for now, but much of this growth is built on guaranteed financing. If credit spreads keep widening, the whole AI capital expenditure cycle could face a sudden squeeze. The real vulnerability lies not in end-user demand, but in whether these financing backstops can sustain the risk.

I still hold the same view: chips, storage and power are no longer the bottleneck — capital is. What will happen to AI hardware names if cheap capital disappears and data center construction grinds to a halt?

[Global Markets Investor @GlobalMktObserv]

🚨AI credit risk is EXPLODING:

Oracle's, $ORCL, 5-year credit default swap (CDS) has surged to more than 200 basis points, its highest level on record, surpassing even the peak seen during the 2008 Global Financial Crisis.

Broadcom's, $AVGO, 5-year CDS has widened by +28 basis points so far in August alone, to a record ~126 basis points.

This comes as Broadcom backstops mega financing packages for the AI buildout, including an existing $35 billion deal with Apollo and Blackstone tied to Anthropic, and a potential further $60 billion in debt under discussion.

Under these structures, Broadcom effectively lends the strength of its own balance sheet to help its customers keep buying chips, meaning it could be forced to honor billions in guarantees during a downturn just as its own earnings come under pressure.

This adds to a broader pattern of deteriorating hyperscaler credit, with CDS levels across the AI ecosystem already back near their all-time wides.

The real AI risk may not be demand. It may be what happens when the companies financing the boom can no longer carry the risk.

[katz @katzspx]

@SemiconductorsX Precisely. KV cache pressure does not directly equal NAND demand, workload optimization is the critical swing factor here for $SNDK.

[katz @katzspx]

The World Robot Games hosted in China has proven that Unitree currently holds little technological moat within China’s robotics sector. What we should focus on is not robot competitions, but a high-performance version of Toy Story. Until humanoid robots can break through the boundaries of the physical world, industrial robots will maintain higher productivity.

[katz @katzspx]

On this topic, there’s only one thing I want to state:

Testing performance is meaningless in the face of mass production.🤡

[WOLF @WOLF_Financial]

OPENAI SAYS ITS JALAPENO CHIP BEAT NVIDIA GB300 IN INFERENCE TESTS

OpenAI says its new Broadcom-built Jalapeno chip outperformed NVIDIA $NVDA GB300 on both throughput per watt and response latency in testing.

Jalapeno runs at roughly 700W and is expected to begin powering OpenAI models later this year, targeting lower-cost and faster inference.

Important caveat: it was not tested against NVIDIA’s newer Vera Rubin platform and is not designed for AI training.

OpenAI says a second-generation chip is already nearing tape-out, while work on a third generation has begun.

[katz @katzspx]

This is a fascinating hardware tradeoff. If OpenAI has to de-spec HBM4 to add more vendors for Jalapeño scaling, how much inference/training performance will they give up to unlock supply? NVIDIA’s Rubin likely built compatibility for multi-vendor HBM from the start, which is a major advantage for its ramp.

[katz @katzspx]

Let me break down why Goldman Sachs projects China will only account for 12% of global HBM TAM by 2028.
▪️The majority of global HBM demand comes from massive AI training clusters run by North American hyperscalers, which creates an extremely large baseline market.
▪️Export restrictions limit China’s access to cutting-edge HBM3E/HBM4. Local infrastructure builders adopt alternatives such as commodity DRAM and KV cache instead of premium HBM.
▪️Most of China’s incremental DRAM demand sits in mainstream DDR memory rather than high-value HBM.
▪️Domestic HBM still faces bottlenecks in process, 3D stacking and advanced packaging, with limited mass-production capacity by 2028.

I think these supply constraints are pushing China to fully support domestic firms to develop the full DRAM industrial chain. Following its IPO, $CXMT is speeding up HBM research and development.

If $CXMT can successfully develop and mass-produce HBM matching the performance of global leaders under these restrictions, it will deliver a devastating blow to the long-standing oligopoly of Samsung, $SKHY and $MU.

[P Equity Research 📰 @pequityresearch]

Goldman Sachs: China Memory

> Massive Overall Growth: Total China DRAM demand is projected to surge from $19,297 million in 2024 to $257,315 million by 2028, achieving a massive 50% CAGR from 2026E to 2028E.

> Explosive HBM Expansion: High Bandwidth Memory (HBM) demand is experiencing hyper-growth, expected to scale at an incredible 188% CAGR (2026E–2028E) to reach $32,105 million by 2028.

> Volume (Shipments): Total shipments are projected to climb from 7,073 mn GB in 2024 to 19,505 mn GB by 2028, with a 44% CAGR expected between 2026E and 2028E.

> Average Selling Price (ASP) Surge: Overall ASPs jump significantly from $2.7/GB in 2024 to a peak of $13.8/GB in 2027 before settling around $13.2/GB in 2028.

Despite this, China's HBM TAM contribution in 2028 is expected to be only 12%.

$MU $DRAM $EWY

[katz @katzspx]

@StockSavvyShay On-device agents are a double-edged sword for $NVDA. It unlocks new edge hardware demand, yet it suppresses recurring cloud token spending over time.

[katz @katzspx]

@SomniaRobotics Exactly, it’s genuinely demanding work！

[katz @katzspx]

Yes, exactly. From the very beginning, China’s stock market was designed primarily as a financing tool for companies—especially as a way to relieve the debt burden of state-owned enterprises. It was never built for value investing.
Things have improved somewhat over the years, but I still don’t fully understand why Beijing is so obsessed with keeping the market stable. On the surface it looks like investor protection, yet the constant intervention ends up dulling market sentiment and price discovery. In my view, the STAR Market may be the only real hope for the future of China’s equity markets.

[katz @katzspx]

It’s less about pure financial investment and more about ecosystem control.

$NVDA doesn’t need to pour fresh cash; its existing equity positions let it align incentives with critical infrastructure players.

The upside is straightforward: stable backup chip supply from $INTC to mitigate GPU capacity bottlenecks, plus early access to next-gen off-world infrastructure from $SPCX, locking in priority for future space-based computing demand without heavy upfront capital outlay.

Yet hidden risks deserve attention.
First, these are unrealized paper gains, vulnerable to broad tech valuation corrections.
Second, deep equity ties may create conflict of interest when competing with Intel in chip markets.
Third, if $SPCX or $INTC faces operational or regulatory setbacks, Nvidia’s balance sheet and strategic roadmap will be dragged down simultaneously.

We’re witnessing how the dominant GPU vendor is building soft power across computing and aerospace, not just hardware market share.

[Leverage Shares ETPs EU @LeverageShares]

@NVIDIA’s equity portfolio quadrupled to $63.4B in Q2 despite making no trades.

$INTC remains its largest holding at $30.0B, while SpaceX—disclosed for the first time following its IPO—accounts for $21.0B.

Other holdings include $CRWV, $COHR, $NOK, $SNPS and $NBIS.