30-yr Optics & Semiconductor Engineer. turned 8x Founder. Now Canadian buy-side investor hunting physical bottlenecks in AI infrastructure. Silicon,
Toronto Canada
Joined December 2022
- Tweets1.5K
- Following235
- Followers1.9K
- Likes724
Pinned Tweet
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.
Power Topology:
10kV AC to 800V DC via Solid-State Transformers is live in pilot, bypassing legacy multi-stage step-down and 48V busbars.
Mechanical Capex:
Eliminating rooftop chillers via free air economization drops build cost to under $3M/MW, compared to $15M/MW in the US.
Execution Velocity:
120MW delivered in 12 months with 4,000 MEP workers on site highlights the industrial disparity.
While US hyperscalers battle substation queues, physical buildout speed dictates cluster scaling. $VRT $NVDA $ETN
The narrative that "NVIDIA guidance relied on smuggling chips" and will trigger a domino collapse collapses under basic financial and supply chain arithmetic:
The Percentage Reality: The widely cited DOJ indictment involves $2.5B of servers over multiple years. Against NVIDIA's $140B+ annualized data center revenue, illicit transshipment represents less than 2% of volume. Direct China revenue is already de-risked to sub-10% in SEC filings.
Who Actually Backstops Neoclouds: NVIDIA provides hardware allocation, not credit guarantees. Neocloud debt facilities ($14.5B CapEx at CoreWeave) are underwritten by private credit (Blackstone, Magnetar) and backstopped by Microsoft take-or-pay capacity contracts. NVIDIA does not guarantee customer debt.
The True Demand Driver: Over 85% of Blackwell and Rubin allocations are absorbed by domestic tier-1 CSPs (Microsoft, Meta, Google, AWS). Anthropic's S-1 reveals $5B in quarterly spot compute absorbed just to meet enterprise inferencing surges.
Enforcing export controls cleanses secondary broker arbitrage without impairing primary hyperscaler cash flows.
$NVDA $MSFT
follow the money:
> (some) nvidia revenue was from smuggling chips
> nvidia/jensen is now compromised
> usgov has leverage to stop this
> future nvidia revenue is at risk
> nvidia backstops every major neocloud
> nvidia backstops every major ai infra partner
> all guidance relied on this chain working
what happens when nvidia revenue slows?
2
2,381
ZeroHedge highlighting CoreWeave's -$7.5B quarterly cash burn ($14.5B CapEx) exposes the structural divide in AI infrastructure financing:
The spike in CapEx reflects physical Blackwell GB200 NVL72 rack deliveries hitting datacenter floors. These deployments are funded via private asset-backed debt facilities (Blackstone, Magnetar) tied directly to Microsoft take-or-pay agreements.
Hardware procurement appears on cash flow statements today, while enterprise cloud billing ramps over 36-month contract terms. Operating cash flow naturally lags capital outlay by 6 to 9 months during cluster bring-up.
Unlike hyperscalers that self-fund CapEx via pristine software operating cash flows, neoclouds carry substantial debt leverage. With interest rates elevated and commercial GPU depreciation set at 3-4 years, any execution slip in power-on timelines directly impairs debt service coverage.
CoreWeave is the levered balance sheet of the AI buildout; hyperscalers remain the unencumbered cash engines.
$NVDA $MSFT
1
301
1/5 A widely circulated institutional memo translated by @jukan05 comparing AI hardware to Chinese real estate(CoreWeave/NVIDIA as Sunac/Vanke, TSMC/ASML as Dehua/Skshu) misdiagnoses balance sheet physics. Residential real estate carries perishable inventory and 85% debt financing. Leading AI semi players self-fund with record free cash flow.
1
2
437
4/5 The bullwhip effect does not break upstream fabs. TSMC and ASML are shielded by non-cancelable take-or-pay LTAs, sub-nanometer ALD tolerances, and CMP planarization limits. A CoWoS-L line configured for Blackwell cannot be arbitrarily scrapped or repurposed without multi-month retooling.
1
81
5/5 Conclusion: This correction reflects a classic mid-cycle multiple compression driven by macro rates, rather than a 2000-style dotcom collapse. As SaaS monetization shifts to project outcomes and Blackwell racks scale past 130kW, value accrues to mission-critical physical interconnects and enterprise workflows.
$NVDA $TSM $ASML $MSFT
78
STH calling AMD's EPYC Venice a "Rome-like cycle" identifies the structural shift in enterprise datacenter silicon: Ecosystem Validation Priority.
In 2019, EPYC Rome forced storage and networking OEMs to qualify platforms on AMD before Intel. With Venice introducing native PCIe Gen 6 and 16-channel memory, SSD controller and 800G NIC vendors must treat Venice as the primary reference design.
Moving to 16-channel MRDIMM breaks the 1 TB/s socket bandwidth barrier. Agentic AI workflows, long-context RAG, and vector databases require massive host memory width to prevent stalling accelerator pipelines.
AMD leverages stable packaging and leading-edge lithography, while Intel Diamond Rapids navigates 18A fab ramp yields.
Platform qualification priority is the ultimate leading indicator of sticky datacenter share.
$AMD $INTC
AMD is kicking ass
ServeTheHome: "Venice is just a big upgrade across the board"
"That means that AMD will benefit from an EPYC 7002 “Rome” like cycle since other vendors (e.g. storage vendors) will have to validate their platforms first on Venice" by @Patrick1Kennedy
servethehome.com/amd-epyc-90…
Insuring supply commitments is standard credit derisking, not hidden debt. Broadcom mandates insurance backstops before booking non-cancelable TSMC 3nm wafer starts. Passing counterparty risk to insurers protects AVGO margins from frontier lab insolvency.
This misreads Anthropic's S-1. The $161B reflects multi-year take-or-pay compute purchase obligations owed BY Anthropic to suppliers, not loans extended by Broadcom. Confusing customer procurement commitments with vendor debt financing reflects basic accounting illiteracy.
Breaking: Michael Burry digs into Broadcom’s deal with Anthropic and finds a problem
Burry found:
• Broadcom promised Anthropic $161B of computer gear
• Broadcom passed some risk to insurance companies
• This is how companies hide debt as a deal gets bigger
Burry also found $AVGO is lending Anthropic $42B on top of the $161B it’s already owed, so he has no plans to buy Anthropic stock when it IPOs next week
1
4
983
Bernstein's note highlights an architectural split between hyperscalers and NVIDIA:
Thermal and Yield Physics:
Continuous-wave (CW) InP lasers suffer rapid MTTF degradation above 65°C, while compute dies run at 85°C+. NPO (Near-Package Optics) places the optical engine adjacent to the package, thermally isolating the optics and preventing laser failures from bricking $30k+ TPU dies.
Serviceability in OCS Fabrics:
Google's Apollo optical circuit switch (OCS) topology requires socketable optical transceivers. True CPO permanently bonds optics to the silicon interposer, creating unacceptable field repair overhead in multi-megawatt TPU pods.
Margin Realities for Innolight and Eoptolink:
Model assumptions of 50% gross margin ignore hyperscaler BOM unbundling. Google directly procures DSP silicon and laser diodes, compressing assembly gross margins to 32-38%.
Even with margin normalization, 12M units represents a multi-billion dollar cash flow inflection for optical packaging leaders.
$GOOGL $AMZN
伯恩斯坦，
英伟达推 CPO 比超大规模云厂激进，后者偏好 NPO；AWS 2027 年在 Trainium 4 上用 CPO，谷歌下了 1200 万只 NPO 模块订单、约 120 至 150 亿美元，中际旭创和新易盛分别拿下 1200 万只 NPO 模块订单的 60% 和 40%！
假设， 1200 万只、单价 1000 至 1250 美元，毛利率按照 50% 推算，
旭创： 511 至 639 亿元收入、147 至 183 亿元净利。
新易盛：341 至 426 亿元收入、136 至 170 亿元净利。
按照目前产能和技术迭代来看，实际是 2027 年 500 至 600 万只、2028 年超千万只！
4
1,056
As next-generation AI racks scale past 250kW towards 500kW, the legacy 48V busbar architecture hits a physical wall across three vectors:
Conduction Loss ($I^2R$): Delivering 300kW at 48V requires over 6,250 Amperes of current. Conduction losses scale with the square of current ($I^2R$). Stepping up to 800VDC slashes current by ~16x and reduces copper resistive heat dissipation by ~270x.
Copper Mass and Volumetric Space: A 6,000A busbar requires solid copper trunks inches thick, crowding out liquid-cooling manifolds. 800VDC reduces busbar conductor cross-sectional area by over 80%.
Wide-Bandgap Power Electronics: Down-converting 800V to sub-1V core VDD shifts the power supply unit (PSU) landscape from legacy silicon MOSFETs to high-voltage Gallium Nitride (GaN) and Silicon Carbide (SiC) modules.
The 2027 timeline aligns with Rubin Ultra deployments, where power delivery joins optical interconnects as a primary cluster bottleneck.
$NVDA $VRT $MPWR $IFNNY
Goldman Sachs states pilot programs are expected in 2027 for 800VDC before broader scaling in 2028.
$NVDA
1
6
870
Anthropic's S-1 discloses the defining micro-economic reality of frontier AI: enterprise inferencing demand is outpacing 18-month hardware planning cycles.
Three institutional takeaways from the $24B short-term compute disclosure:
The Spot Premium Tax: Anthropic absorbed $5.0B in short-term compute in Q3 alone. The $5.4B incremental revenue was served entirely on emergency spot capacity, driving revenue per compute dollar from $2.08 to $1.72 and causing a temporary $570M operating loss.
The Arbitrage Bridge: Paying exorbitant hourly premiums to secondary operators (including SpaceX GPU clusters) represents an operational buffer while awaiting contracted capacity. Demand is outrunning supply chain delivery schedules.
The Anthropic Direct Inflection: With $24B in short-term commitments out of $70.7B due over 12 months, bringing dedicated leases and Google TPU clusters online in 2027 will replace $3-4/hr spot rentals with sub-$1.50/hr contracted rates, driving substantial gross margin expansion.
Operating losses reflect extreme demand elasticity
What Anthropic's S-1 discloses about its expensive short term chip rentals (aimed at SpaceX)
- Most compute is secured 12 to 18 months in advance under multiyear contracts. When customer or research demand exceeds that capacity, the company buys short term compute, which "currently" costs "significantly more" than comparable compute under multiyear contracts.
- The scale was $5.0 billion in Q3 and $6.5 billion over the first nine months.
- Operating income of $970 million in Q2 turned into an operating loss of $570 million in Q3. The company attributes this to $5.0 billion of short term compute and $2.0 billion of noncash donation matching charges.
- Use of short term compute will keep increasing in the near term, and operating margin will be compressed as a result.
- As of the end of September, minimum payments under short term contracts (cancelable without penalty within one year, or with a remaining term of one year or less) were about $17.4 billion. About $6.6 billion more has been signed since October.
- Once Anthropic Direct, under which the company leases chips directly and hosts them in dedicated data centers, comes online, the company will replace the short term compute bought during demand surges with cheaper long term compute.
Interpretation:
- Excluding short term capacity, Q3 compute cost was $4.86 billion, below the Q2 total of $5.56 billion. Long term contracted capacity barely grew in Q3. In effect, the $5.4 billion increase in revenue was served with short term capacity.
- Revenue per dollar of compute fell from $2.08 in Q2 to $1.72 in Q3. The company also notes that this metric declined from the prior quarter.
- The $17.4 billion short term contract balance plus the $6.6 billion of new contracts comes to $24.0 billion, about one third of the $70.7 billion due over the next 12 months. More than the Q3 pace ($5.0 billion a quarter) is already under contract.
Question: Will Anthropic keep paying SpaceX a huge premium after March 2027?
Source: Anthropic’s S-1 filing
Every macro pullback, Taiwanese supply-chain media publishes the same "NVIDIA cuts TSMC orders" rumor. In reality, physical wafer fabs operate under rigid manufacturing constraints:
Architecture Migration vs Net Cuts: Shifting wafer starts from 4NP (Hopper H200) to N3P (Blackwell B200 / Rubin R100) is scheduled node transition. Reallocating production lines does not reduce aggregate wafer absorption.
Packaging Non-Fungibility: CoWoS-L lines for Blackwell run a 3.3x reticle size with local silicon interconnect (LSI) bridges. This tooling cannot dynamically transfer to AMD MI350 or custom ASICs (Alchip, MediaTek) without a multi-month substrate retooling cycle.
Hyperscaler Backlog Realities: Microsoft, Meta, Google, and Amazon maintain multi-quarter purchase commitments. With rack lead times still exceeding 24 weeks, canceling wafer starts makes zero commercial or operational sense.
$NVDA $TSM $AMD
A Taiwanese media report says NVIDIA may cut its orders with TSMC. If that happens, AMD and ASIC chipmakers could pick up some of the wafer capacity currently allocated to NVIDIA.
I’ve heard similar rumors. I’ll let you know as soon as I can confirm them.
ctee.com.tw/news/20261110700…
ZeroHedge framing chipmaker CDS widening as an AI credit collapse misreads fixed income market microstructure:
Multi-strategy funds cannot easily short physical wafer starts or CoWoS allocations. Buying chipmaker CDS serves as a liquid synthetic proxy to hedge long equity books against geopolitical export controls.
Hyperscalers (Microsoft, Meta, Google) run diversified, recurring software and advertising cash engines. Chipmakers carry working capital absorption, financing billions in non-cancelable wafer start commitments at TSMC.
A 156 bps CDS spread represents an annual insurance cost of 1.56%. That is mild investment-grade credit tightening. True distressed credit trades above 600 bps.
With $NVDA holding tens of billions in net cash and record free cash flow, this is balance sheet inventory duration hedging, not a solvency crisis.
"Risk Is Migrating": Chip CDS Blows Out Wider Than Hyperscalers For The First Time, As AI Default Swap Volumes Hit Record zerohedge.com/markets/risk-m…
Shay's bullish thesis on $AMD captures two structural tailwinds while overlooking an architectural shift:
Software Friction Neutralized:
LLM coding agents and OpenAI Triton have largely solved ROCm porting. PyTorch workloads compile across CDNA with minimal manual kernel tuning.
Agentic CPU Demand:
Long-context agent orchestration, tool-calling sandboxes, and JSON parsing create an unexpected demand cycle for high-core-count EPYC Turin nodes.
The Real Architectural Gap: Custom silicon customers (Google TPU, AWS Trainium) increasingly bypass external x86 host CPUs for in-house ARM cores (Axion, Graviton) to eliminate PCIe power penalties.
Scale-Up Fabric Physics: The battleground is no longer CUDA vs ROCm, but scale-up domain size. AMD must expand Infinity Fabric beyond 8-GPU nodes to match NVIDIA's 72-GPU coherent NVLink copper domain.
EPYC and CDNA are executing, but the moat is physical interconnect topology.
$AMD $NVDA $ARM
I could not be more bullish on $AMD as a business right now because its GPUs are finally moving into multi-gigawatt deployments just as agentic AI is creating an entirely new growth cycle for EPYC CPUs.
What makes this combination so interesting is that even customers building their own AI chips still need host CPUs while coding agents and ROCm are making AMD's GPU platform way easier to adopt.
Sell-side models projecting 5GW of SpaceX space compute by 2028 via 10k "Kessel" Rubin racks fail basic orbital physics across three dimensions:
Radiative Thermal Physics: In vacuum, heat can only dissipate via thermal radiation. Under the Stefan-Boltzmann law at 70°C, emissivity yields ~400 W/m². Radiating 5GW requires 12.5 square kilometers of deployable radiator surface (~1,250 m² per rack).
Launch Mass Constraints: A compact 500kW liquid-to-radiator rack with 4,000W TDP GPUs, solar wings, and pumps masses ~8 metric tons. 10k racks represent 80,000 tons in LEO. That requires over 530 flawless Starship launches in an 18-month window solely for compute payload.
Radiation Hardening on TSMC 3nm: Commercial leading-edge logic dies are vulnerable to single-event upsets (SEU) and latchups from cosmic rays. Hardening or heavy shielding multiplies structural mass.
Space compute will scale, but 5GW by 2028 is paper engineering.
$NVDA
Morgan Stanley highlighting memory multiple compression marks a critical sentiment inflection.
Three structural shifts separate this cycle from historical boom-bust memory traps:
1. Buybacks Over Greenfield Floods:
Previous upturns saw suppliers burn operating cash flow on unhedged cleanrooms. Today, Micron ($MU) and SK Hynix ($SKHY) channel structural HBM free cash flow into retiring shares at compressed valuation multiples.
2. The Infrastructure Delay Shield:
Land access, substation lead times, and cleanroom shell delays physically cap wafer fab velocity. Supply cannot ramp fast enough to overshoot enterprise demand before 2027.
3. Pricing Architecture:
Multi-quarter advance purchase commitments for HBM and server DDR5 decouple cash generation from spot commodity volatility.
Multiple compression against record structural FCF yield creates classic buy-side asymmetry.
$MU $SKHY $DRAM
Morgan Stanley on memory:
"Memory risk reward starting to look very favorable even with land/power/shell risks. For us, the multiple compression in memory is increasingly offering very compelling upside relative to those risks, and FCF earned during the upturn is being predominantly used to buy back large quantities of stock."
$MU $DRAM $EWY $SKHY $SNDK
$SITM content growth is driven by physical rack density. In 100kW+ liquid-cooled clusters, acoustic and mechanical vibrations induce severe phase jitter in traditional quartz.
Silicon MEMS solves three physical constraints as clusters scale to 224G PAM4:
1. Resonators deliver sub-0.1 ppb/g g-sensitivity, preventing link retrains when coolant pumps cycle.
2. Micron-scale MEMS dies integrate directly into optical engine substrates where quartz cans cannot fit.
3. Tighter femtosecond jitter windows at PCIe Gen 6/7 make precision timing mandatory at every retimer and switch.
The key risk to monitor is customer dual-sourcing as timing becomes a non-trivial line item in cluster BOMs.
$SITM is my largest Growth Portfolio position because SiTime is essentially clock inside AI keeping GPUs, CPUs, switches, memory and optical links synchronized so expensive compute doesn't go to waste.
Where this gets really interesting is that inference can drive 2-3x more timing content per CPU server while near-package optics can roughly double timing content, co-packaged optics can triple it and coherent optics bring much higher selling prices.
So the beauty is that SiTime doesn't just benefit from more AI servers being built but captures more revenue from every system as architectures get more complex and it doesn't matter whether $NVDA keeps dominating, $AMD takes share, $ARM keeps growing or hyperscalers build more custom silicon because every one of those architectures still needs timing.
Culper Research (@CulperResearch) published a forensic 17-part short thesis on Nvidia ($NVDA) detailing a Southeast Asian GPU diversion pipeline involving Alibaba, Megaspeed, and Aivres.
While the corporate registry forensics are detailed, the short thesis commits a fatal semiconductor economics error:
The "$30B Revenue Air Gap" Fallacy: Culper assumes that freezing Southeast Asian smuggling wipes out $30B in Nvidia revenue. This treats Nvidia as demand-constrained. In reality, Nvidia is strictly CoWoS-allocation-constrained. TSMC packaging lines are 100% committed through late 2026.
The Margin Expansion Paradox: Diverted H200 chips carry regulatory discounts and distributor friction. If Washington closes the Singapore/Malaysia loophole, Nvidia simply reallocates those scarce CoWoS packaging slots to backlogged US hyperscalers waiting 12 months for GB200 and GB300 NVL72 racks. Reallocation increases blended ASP and expands gross margins.
Culper mapped a smuggling corridor, but mispriced the physics of wafer fungibility.
$NVDA $BABA $SMCI
1
8
2,543