The Indium Phosphide Bottleneck: Why Silicon Photonics Rescued Pluggable Optics—and Triggered a Global Laser Wafer Crisis
Inside the 7-stage photonic supply chain, China's MOFCOM controls & the 2.1x CW wafer yield salvation inside 1.6T transceivers.
Calling Max Overweight after an institutional deleveraging panic is textbook cycle mechanics. Hedge funds were forced -11% underweight tech on factor stops, leaving no marginal sellers. Recommending Japan and Europe reflects capex rotating into global power and semicap hardware.
At 15x fwd P/E for MSCI ACWI, equity risk premiums absorb higher rates far better than 30x Mag-7 multiples. The call is a duration rebalance: extracting AI productivity gains through diversified industrials and optical suppliers while avoiding pure hyperscaler capex fatigue.
Morgan Stanley pitching Europe and Japan alongside AI is an infrastructure tell. As US hyperscalers hit domestic grid walls, capex rotates into global power equipment, thermal distribution, and Asian semicap. It is not abandoning AI, but funding the physical supply chain.
Morgan Stanley recommends “max overweight” on global stocks, citing an intact AI capital spending cycle that will extend resilient economic and earnings growth into next year bloomberg.com/news/articles/…
1
1
175
Seeing ESMT at 1.2x P/E shocks retail, but cyclical memory always trades at trough multiples at peak earnings. Q3's 77% margin reflects temporary DDR3 panic hoarding from HBM wafer crowd-out. The market values it at liquidation because spot windfalls fade fast. 3006.TW
I'm actually very impressed with market-chan. If we look at ESMT numbers today:
> $1.58B MC
> $747M net cash, $274.7M net inventory from Q3
> $349.5M in Q3 free cash flow
Adjusted EV: $559.3M
> Q3 net income: $327M (1.2x annualized P/E)
> October net income (likely another +$96M added to balance sheet)
- "DDR2/DDR3 supply to remain constrained into 2027"
- "ASPs to rise Q4" (consumer demand weakening)
- "2027 dividend payout ratio could be 30–50% of earnings"
So you're valuing the adj. operating business at $559M, despite them making ~$300m net income a quarter.
Even the AI models I've double checked it with looked at these numbers with disbelief.
191
The 8-week optical selloff confuses commodity module assembly with upstream component moats. CPO saves 3kW per switch to free 45MW for GPUs, while optical pooling cuts HBM4 by 30%. True alpha accrues to CW lasers and InP substrates, not dilutive packaging assemblers.
In 1GW power-limited campuses, CPO is an efficiency arbitrage. Saving 3kW per switch frees 45MW of network overhead, adding headroom for 3,100+ GPUs without new grid permits. Efficiency gains decouple optical component capex from aggregate hyperscaler budget cuts. $COHR $LITE
True, I've looked at biotech, physical AI, cybersecurity, space, and others... But I still have conviction that optics is the most compelling opportunity.
If you look at Elon Musk $SPCX (Mesh acquisition), $MRVL (Lumentum acqusition), Samsung Electronics photonics foundry push...
SK Hynix CPO push + PIC IP acqusitions... it's endless.
Private VCs are still heavily pouring funds into upcoming optical companies.
Supply procurement is accelerating... from $SOI selling $1.2B in agreements, in a single quarter (from $25M revenue/quarter). To $COHR signing $20B in LTAs, to $LITE selling out of 2029 capacity.
Actions about where the industry is heading is the biggest signal over valuations crashing.
Even if you stress test aggregate capex slowdown or accelerator flood vs. grid constraint/permitting downside:
- There's an economic incentive to improve GPU efficiency and power utilization. (eg. Lambda's CPO data suggests enough power savings to support ~4.5% more GPUs)... for power constrained DCs.
- There's an economic incentive to reduce expensive HBM4 requirements in certain cases (eg. $MRVL estimating up to 30% savings through optical memory pooling)... for memory constrained inference.
While for economic upside in these companies:
-> Optical networking, and their upstream materials like substrates and lasers represent a relatively small fraction of total AI cluster BOM.
-> At the same time, optical content goes up...while selling prices of that content rise nonlinearly as well (lasers 10x+ selling price from pluggable to CPO.)
In terms of the future winners:
- Many low end players eg. 400G/800G, module assemblers etc. don't capture the same scale in economics from the NPO/CPO transitions... compared to laser suppliers, fiber array suppliers, modulator suppliers, or companies in the ELS/optical engine stacks.
- Tracking capacity expansion is also important... but US/China supply chain birufration will improve scarcity economics of Western players (eg. China optical transceiver bans past 3.2T)
But even if aggregate AI capex slows, there's a strong economic incentive to accelerate optical deployments and to make every iota of compute more economically valuable.
Visibility keeps increasing, demand keeps increasing, strategic value keeps increasing, while valuations crashed. NFA, but I'd personally, argue this is an opportunity.
Extremely bullish on industry transition to light.
1
5
640
Dismissing Kessel misses unit economics. On a $3.2M NVL72 rack, Nvidia captures $2.4M gross profit. Selling bare silicon at $1.8M drops profit to $1.44M even at 80% margin. Stripping $1M in rack gross profit per cluster commoditizes Nvidia's system integration moat.
The risk is not SpaceX buying chips, but the blueprint it establishes. If an 84-layer midplane proves customers can bypass proprietary NVLink cable assemblies and power shelves, other hyperscalers will unbundle too. System commoditization compresses blended ASPs. $NVDA
There’s a lot of careless analysis dismissing Nvidia advantages.
Nvidia gets the highest margins on the GPU themselves. They capture more market by delivering more content in the rack. But their margins on the GPU are still very high.
Custom silicon for hyperscalers is a small piece of the market, and until they get more efficient in tokens per watt, Nvidia still crushes.
The idea that it’s a problem for Nvidia that SpaceX is developing AI satellites around Nvidia GPU’s is bizarre. This increases the market for Nvidia, a positive.
The biggest threats to Nvidia are losing key engineers, and losing relationships with their supply chain.
1
494
The 879k chip leak reflects the divide between DoD and Commerce. Pentagon 1260H listings do not trigger automatic EAR export bans; shipments held formal BIS licenses. Lawmakers will weaponize this data in the Dec 2 hearing to revoke all remaining exemptions.
This leak is the evidentiary dossier for the December 2 Congressional hearing with Nvidia's General Counsel Tim Teter. Lawmakers will weaponize the 879k figure to grill Howard Lutnick's Commerce nomination and push for a complete embargo on downgraded chips. $NVDA $AMD
Revealed: Trump Admin Cleared 879,868 Nvidia, AMD AI Chips For China, A Third For Pentagon-Listed Military Companies zerohedge.com/geopolitical/r…
1
1
508
Hardware brokers actively shopping "orphaned B300s" confirms that the AI compute bottleneck has inverted from silicon to energized power.
The market fixates on GPU availability, yet secondary channel data reveals physical infrastructure realities:
Powered Shell Scarcity: Thousands of HGX B300 servers originally built for Chinese offtake are stranded. Plugging 100kW liquid-cooled racks into commercial colocation is impossible without high-density power distribution. In the US, utility substation queues stretch 4 to 5 years. Hardware cannot outrun the grid.
The Regulatory Cat-and-Mouse: Brokers using the Wayback Machine to screen out "AI slop" neocloud sites illustrates how aggressively Chinese buyers attempted offshore proxy arbitrage. As BIS closes front-company loopholes, unallocated servers face fire-sale liquidations into Western spot markets, depressing hourly rental rates.
Capital Structure Polarization: Retail lamenting IREN's stock weakness while data center peers rally misses financing mechanics. CoreWeave locked in $25B in non-dilutive DDTL, and Nebius self-funds via 100%+ customer prepayments. IREN, historically reliant on ATM equity dilution, faces a heavy multiple discount as it stretches Childress capex.
Silicon is turning into a commodity of fire sales. Energized megawatts remain the absolute sovereign monopoly. $NVDA $IREN $CRWV
3
6
1,253
Mark Gurman confirming Apple's autonomous agent debunks the 'partner with Google' thesis. Search returns web links; proactive agents execute OS actions across 2.2B devices. Outsourcing the agent surrenders the operating system. Apple must own the native action runtime.
Apple will not copy Meta's commercial GPU CapEx frenzy.
John Ternus driving the AI push reflects Apple's silicon co-design playbook. $AAPL
$AAPL is now developing its own AI agent offerings to compete vs. $ANT, Meta Muse, Google Gemini, and OpenAI.
Given on-device inference isn't quite ready yet for complex workflows...
+ Apple isn't really known for doing much in the AI race
This is pretty positive for aggregate inference demand + infra capex implications (as seen with Meta Muse supply procurement).
It also kinda follows rumors they were looking to acquire private accelerator companies for their own inference buildout.
440
Apple developing a proactive agent leverages its 2.2B active device moat. While OpenAI and Anthropic build frontier models, they remain sandboxed apps. Apple controls the system daemons, App Intents, and M-series Private Cloud Compute needed for autonomous action. $AAPL
SemiAnalysis reporting that SpaceX is developing a custom rack architecture codenamed "Kessel" reveals the secondary front threatening Nvidia's valuation.
The consensus fixates entirely on custom silicon substitution, noting that Google, Microsoft, Meta, Amazon, OpenAI, and Anthropic are building in-house chips. SpaceX remains the lone non-silicon customer. Yet treating SpaceX as a standard turnkey buyer misreads system-level hardware economics:
System Margin Disintermediation: Nvidia's highest incremental margins on NVL72 come from rack-scale integration, charging hundreds of thousands of dollars in markups on proprietary NVLink copper cartridges, power shelves, and switch trays. SpaceX buying bare chip packages to build Kessel completely unbundles these peripheral profits.
The 84-Layer M10 Midplane: Kessel discards Nvidia's internal NVLink copper cables entirely, adopting an 84-layer passive PCB midplane using ultra-low-loss M10 laminates. Squeezing compute and switch trays back-to-back is an aerospace engineering necessity: flexible copper cable harnesses cannot survive Starship launch vibrations or enable efficient vacuum thermal conduction.
Simplifying Kyber: While Nvidia struggled to yield its 156-layer M9 Kyber architecture, SpaceX optimized the midplane down to 84 layers of M10 to balance 224Gbps PAM4 signal integrity with commercial board fab yields.
Nvidia faces a bilateral squeeze: six hyperscalers are commoditizing its silicon with ASICs, while SpaceX is commoditizing its system-level rack integration. The full-stack monopoly is dissolving from both ends. $NVDA
Hunterbrook shorting AMD on China channel scrutiny exposes the fallacy of the safe-haven trade. US fabless chipmakers share identical Asian distributor and ODM channels. As BIS tightens export enforcement, regulatory scrutiny is systemic, not company-specific.
Repackaging distributor gray-market leaks into a bespoke AMD scandal generates short-seller volatility. $AMD $NVDA
Investors treated $AMD as the winner of Nvidia’s China crisis.
The problem: AMD has a China problem of its own.
Our investigation:
Readers added context they thought people might want to know
Hunterbrook reports AMD TurinPI firmware in post-Mar 2025 updates from Entity List-listed IEIT Systems (added 3/25/2025, denial policy) and linked firms. Report states it does not establish AMD violation or how firmware was obtained. Hunterbrook Capital is shorting $AMD. govinfo.gov/content/pkg/FR…
Seaport warning that wiping out humanity may lead to a serious decline in earnings captures late-cycle research absurdity.
Reducing human extinction to earnings degradation highlights the absurdity of sell-side DCF modeling. Modeling cancer cures as an upside valuation revision while treating existential wipeout as margin compression proves Wall Street will financialize anything. Peak cycle.
Beyond the viral jokes, Seaport's real warning is off-balance sheet debt. Multi-year compute lease commitments act as shadow debt. If Anthropic fails to generate cash flow to service these unrecorded liabilities, financial leverage will bite long before AGI arrives. $ANT
... on the other hand "If Anthropic cures
cancer it would greatly increase its value beyond our current valuation estimate."
1
529
Needham modeling 150% growth and 35% operating margins is standard post-IPO cheerleading. The $150 target anchors retail sentiment. But sustaining 35% margins while funding $10B+ frontier training clusters requires unprecedented operating leverage from software tokens.
Comparing Anthropic to Google and Microsoft ignores infrastructure ownership. Hyperscalers own the land, power, and custom silicon. Anthropic is a high-volume tenant paying compute lease margins to AWS and GCP. Tenant margins cannot structurally match landlord moats. $ANT
JUST IN: Needham starts coverage of Anthropic, $ANT, with a Buy rating and a $150 price target, about 81% above its $83 IPO price.
Needham expects Anthropic's revenue to grow more than 150% in fiscal 2027, with operating margins above 35% and gross margins above 90%.
It compares Anthropic's long-term potential to Amazon, Google and Microsoft.
Anthropic researchers predicting total human unemployment by 2027 is textbook capability marketing ahead of its IPO. Framing software automation as inevitable doom creates investor FOMO while erecting regulatory moats against open-source rivals. Watch the economics, not PR.
BREAKING: An Anthropic researcher says AI will likely leave “pretty much all humans on Earth” permanently unemployable and dependent on government welfare.
Tao Lin, who works on safeguards for Anthropic’s Claude models, said in an interview with Palisade Research that he expects “broadly superhuman AI,” as good as or better than the best human expert in almost every field, by around the end of 2027.
He said people would eventually “need to strictly rely on government welfare to sustain themselves.” But he doesn’t expect mass unemployment by 2027. He sees a gradual transition in which workers first lose their current roles and move into temporary jobs, partly because “there aren’t enough robots yet.”
“People should be angry at me and everyone else developing AI,” Lin said.
2
365
InP backlog surging to $500M and Photonics-SOI LTAs reaching $1.2B validates 1.6T optics. 200G per lane architectures demand massive Indium Phosphide CW laser arrays and thick-BOX SOI wafers. Upstream substrate scarcity protects Coherent's internal laser margins.
Optical BOM content per cluster scales exponentially into 1.6T. Scale-across topologies require higher transceiver-to-GPU ratios to offset copper reach limits. Coherent retiring $2B in equity confirms structural optical demand over secondary market spot noise. $COHR
That’s the signal of confidence I like to see: $COHR issues a new $2B share buyback, through open market purchases.
Optical industry looks very undervalued right now, especially as it reaches the inflection of the next supercycle.
We’re already seeing photonics-SOI substrate players go from $25M revenue -> $1.2B LTAs signed in a quarter.
Or $100M backlog -> $500M in a quarter with InP substrates.
And we’ll see acceleration for lasers, modules, and the optical supply chain next.
I’m sure Coherent finds its own valuation very low now too, especially after signing $20B LTAs recently.
335
Google using New Glenn for orbital TPU racks reflects terrestrial grid congestion. In vacuum, radiating 80kW requires 250+ square meters of cooling panels. Free-space laser bandwidth restricts space compute to inference, trading utility delays for thermal physics. $GOOGL
Hedge funds sitting 11% underweight Tech against a 29% benchmark weight creates a textbook short squeeze setup. After record factor liquidations washed out momentum, pods face massive negative convexity if mega-caps rebound. The path of least resistance is up.
Goldman's trading desk playbook is classic. First, prime brokerage processes forced pod-shop factor de-grossing. Then, sales trading flags an 11% underweight to ignite a short squeeze. Wall Street monetizes the panic on the way down and the squeeze on the way up. $NVDA $QQQ
Hedge Funds Are Massively Underweight Tech On Anthropic Day, Goldman's Privorotsky Smells A Squeeze zerohedge.com/markets/hedge-…
1
2
677
FT's $602B revenue calculation conflates a multi-year $1T compute pipeline with a single-year asset base. Amortized over 5-7 years, the required hurdle rate drops dramatically. More importantly, hitting an $81B run-rate shows agents are capturing labor payroll budgets.
Goldman projecting $600B by 2030 reflects lead underwriter incentives for Anthropic's upcoming IPO. Circular funding from Amazon and Google subsidizes the compute lease bill today. Long-term 40% operating margins will require genuine software operating leverage. $AMZN $GOOG
Anthropic would need about $602 billion in annual revenue to earn a 20% return on its roughly $1 trillion in compute commitments, per FT
The math: a 20% return on $1 trillion is $200 billion in after-tax profit. At a 17% tax rate, that’s about $241 billion before tax, and at a 40% operating margin, it takes about $602 billion in revenue.
That’s more than Alphabet’s $471 billion or Apple’s $478 billion in annual revenue. It’s about two-thirds of the $908 billion earned by all 76 software and IT services companies in the S&P 1500 combined.
For comparison, Anthropic’s annualized revenue run rate hit $81 billion at the end of October. Goldman, the lead underwriter, projects revenue of more than $600 billion by 2030.
The debate around Nvidia's $203B commitment jump in its 10-Q exposes two major market misunderstandings.
First, replies conflate hard upstream supply commitments with contingent downstream neocloud leases:
- The $390B supply chain commitments ($203B for FY28) are irrevocable take-or-pay contracts with TSMC, SK Hynix, and Micron. Upstream foundries do not accept customer financing conditions before dedicating cleanroom capacity.
- The financing contingency cited in the filing applies strictly to $74B in uncommenced data center leases. Nvidia structurally insulated its balance sheet by making its facility guarantees contingent on neoclouds securing capital.
Second, using $203B in commitments to project $700B+ in FY28 revenue at 72% gross margin fundamentally misreads semiconductor working capital.
HBM4 and CoWoS allocations require cash prepayments and capacity reservations 18 to 24 months in advance. The $203B figure tracks multi-year component pipelines, not a single-year cost-of-goods-sold figure.
Nvidia's disclosure proves hyperscaler baseline demand remains locked, while establishing a contractual firewall against debt-laden neocloud defaults. Balance sheet mechanics separate alpha from headline noise. $NVDA























