Own the bottlenecks, not the apps.
Finding moats in compute, photonics, memory, energy, grid, digital scarcity.
@joinautopilot portfolio managed by @haztan
Underneath every AI app is a supply chain that can't expand fast enough: compute, HBM, packaging, photonics, EUV, power, cooling, grid.
We invest in these bottlenecks and identified four clear sleeves:
• AI Semiconductors 57%
• Energy 17%
• Grid & Thermal 14%
• Digital Scarcity 12%
Full strategy and live portfolio on @joinautopilot ⬇️
https://t.co/02QOL7h7hn
What would actually trigger an exit from one of these positions?
For AMKR: if AI chip packaging moves entirely in-house to TSMC at scale with no room for OSAT participation. Not happening yet.
For CEG: if nuclear plant closures accelerate faster than new PPAs are signed. Worth watching but not the current trajectory.
For IBIT: if Bitcoin loses the neutrality properties that make the settlement thesis work. Protocol-level change. Extremely unlikely but the watch condition.
Positions worth watching more closely:
AMKR 8%: CoWoS capacity expansion at TSMC requires thesis refinement from 'acute spillover' to 'structural packaging market growth.' Still passes the scorecard but conviction is slightly lower.
SNPS 2% + TSEM 2%: watch positions remain appropriately small. No change warranted.
VST 8%: merchant power thesis intact. Grid tightness in ERCOT and PJM remains the driver.
No exits triggered. All 13 positions pass.
Own the bottlenecks, not the apps.
https://t.co/02QOL7h7hn
Two months in, one position required a thesis refinement, zero exits triggered.
Every holding gets re-scored against the chokepoint framework. No grading on a curve.
The honest mid-cycle check, position by position. Thread.
Positions where conviction has increased:
MU 12%: geopolitical tailwind for U.S.-origin HBM adds to fundamental supply scarcity thesis.
CEG 9%: hyperscaler PPA deal pipeline expanding. Existing fleet thesis compounding.
COHR 11%: CPO transition timeline becoming clearer. Both pluggable and co-packaged exposure intact.
Positions that remain unchanged:
NVDA 15%, ASML 7%, VRT 6%, ETN 5%, PWR 3%, IBIT 12%: all pass the scorecard on every dimension. No reason to reduce.
Co-packaged optics doesn't make pluggable transceivers obsolete overnight.
The transition will run in parallel for years as different cluster sizes and use cases adopt CPO at different rates.
COHR wins in both regimes: they supply the lasers for pluggable today and are positioned for CPO as it scales. That's the durable part of the 11% weight.
COHR makes the InP lasers and optical components that go into both current pluggable transceivers and next-generation co-packaged designs.
TSEM has silicon photonics foundry capability - the integration layer that allows optical components to be manufactured on silicon substrates compatible with chip packaging processes.
Both positions have an explicit line into the transition that's happening over the next 2-4 years.
Own the bottlenecks, not the apps.
https://t.co/02QOL7h7hn
At 200,000+ GPU clusters, copper interconnects physically cannot move enough data between nodes.
The fix is co-packaged optics, and almost nobody outside technical circles is talking about it.
COHR and TSEM sit directly in its path. Thread.
Today's AI clusters connect GPUs through electrical copper interconnects and pluggable optical transceivers at the network edge.
Co-packaged optics moves the optical components directly onto the switch or accelerator package - eliminating the electrical-to-optical conversion step that currently limits bandwidth density and increases power consumption.
At 200,000+ GPU clusters, co-packaged optics isn't optional. It's the only way to move enough data between compute nodes.
The AMKR thesis refinement for 2026 and beyond:
The original thesis was spillover from TSMC CoWoS tightness. The evolved thesis is AI packaging demand growth outpacing available capacity across all tiers.
Both can be true simultaneously. The second is more durable.
The key question for the AMKR thesis: as TSMC CoWoS capacity expands, does spillover demand to OSATs like Amkor actually increase?
The answer is nuanced. TSMC's captive packaging serves the most leading-edge designs. AMKR captures the second tier and adjacent packaging work that TSMC doesn't prioritize.
As the overall AI packaging market grows, AMKR's addressable volume grows with it. The thesis doesn't require TSMC to remain constrained. It requires AI packaging demand to stay large.
Own the bottlenecks, not the apps.
https://t.co/02QOL7h7hn
"Advanced packaging is constrained" has been the consensus for two years.
By mid-2026, that's only half true. TSMC added real CoWoS capacity. The AMKR thesis survives anyway, and here's why. Thread.
CoWoS - Chip on Wafer on Substrate - is TSMC's proprietary advanced packaging technology used in NVIDIA's H100 and successor chips.
Through 2024 and 2025, CoWoS was genuinely supply-constrained. TSMC was expanding aggressively but demand from NVDA alone was absorbing nearly all available capacity.
By mid-2026, TSMC has added meaningful CoWoS capacity. The acute shortage has eased somewhat. But demand is also growing - and the constraint hasn't disappeared.
A concentrated thematic portfolio with genuine Bitcoin conviction holds 12%, not 30%.
The reason: thesis uncertainty compounds over time. Bitcoin's settlement layer adoption may take 10+ years to fully materialize.
In the meantime, the AI semiconductor and energy positions are much closer to present-tense cash flows. The 12% buys the option without betting the thesis on it.
The uncertainty worth naming honestly:
This thesis may be right about Bitcoin's properties and wrong about adoption timing. A 12% weight reflects real conviction that the thesis is directionally correct while acknowledging that the timeline is long and uncertain.
Digital scarcity is a real phenomenon. Bitcoin is the leading candidate for the reserve role. The AI buildout is the clearest near-term driver of demand for exactly that role.
That's why it belongs in this portfolio.
Own the bottlenecks, not the apps.
https://t.co/02QOL7h7hn
This thesis may be right about Bitcoin's properties and wrong about adoption timing.
That sentence is why the sleeve is 12% and not 30%.
The complete digital scarcity argument, one week of analysis condensed. Thread.
The argument in three steps:
1. AI agents are already transacting and the volume will grow orders of magnitude.
2. Existing payment rails are structurally unsuited for stateless autonomous machine-to-machine commerce at global scale.
3. Bitcoin is the only monetary asset with the combination of properties - neutral, scarce, permissionless, proven - that a genuine global machine settlement layer requires.
Volatility is the most intuitive objection but arguably the least disqualifying at the settlement layer.
If an AI agent is settling a transaction in Bitcoin and immediately converting the received amount to whatever it needs, the price at settlement is just a momentary state.
The unit of account problem matters more for long-term contracts than for micropayment settlement. That distinction is often skipped.
Reply with A, B, C, or D.
This is an honest exploration of the thesis, not a promotional exercise. The Bitcoin sleeve is 12% because these risks are real and sized accordingly.
Own the bottlenecks, not the apps.
https://t.co/RexuN9tRue
The Bitcoin sleeve is 12%, not 25%, for a reason: the risks to the settlement thesis are real.
Volatility, regulation, competition, or adoption friction. Which one actually kills it?
Your call, A through D. Thread.
A. Price volatility. Bitcoin's price swings make it unreliable as a unit of account for commercial transactions.
B. Regulatory risk. Governments will intervene before autonomous AI systems can hold and transact Bitcoin at scale.
C. Technical competition. A faster, cheaper, purpose-built alternative emerges and gets adopted by the dominant AI platforms.
D. Adoption friction. The AI development community doesn't converge on Bitcoin specifically - agents use whatever the platforms they run on support.