Value Chain Koala

Mapping macro trends into value chains, bottlenecks, and tickers

[Value Chain Koala @valuechainkoala]

I turn messy industries into a clean structure.

Every week:
• the mega catalyst shaping an industry
• the value chain: who gets paid, who gets squeezed
• the bottleneck: where pricing power concentrates
• the tickers and catalysts: the practical takeaway

Research you can actually use.

Full research: link in bio

[Value Chain Koala @valuechainkoala]

Every time $NVDA drops an open-weights model, the conversation is about the model.

The money is in the infrastructure.

550 billion parameters. Every enterprise that wants a custom version fine-tunes it on its own data.

That takes a real cluster, not a consumer GPU, not a cloud API.

That demand flows to $IREN $CRWV $NBIS.

Open weights + enterprise fine-tuning = structural GPU rental demand that doesn't get priced at launch.

Watch neo-cloud utilization in earnings. That's where Nemotron shows up.

[Value Chain Koala @valuechainkoala]

China put InP under export control 15 months ago $AXTI actions since then are your real tell.

They’ve grown backlog to record levels, say capacity is below demand, and just greenlit a $100M Beijing expansion, which you don’t do if you think permits will choke the business.

That capex choice is management explicitly betting they can keep shipping at scale out of China, not signaling an existential block.

[Value Chain Koala @valuechainkoala]

China put InP under export control 15 months ago $AXTI actions since then are your real tell.

They’ve grown backlog to record levels, say capacity is below demand, and just greenlit a $100M Beijing expansion, which you don’t do if you think permits will choke the business.

That capex choice is management explicitly betting they can keep shipping at scale out of China, not signaling an existential block.

[Value Chain Koala @valuechainkoala]

China put InP under export control 15 months ago $AXTI actions since then are your real tell.

They’ve grown backlog to record levels, say capacity is below demand, and just greenlit a $100M Beijing expansion, which you don’t do if you think permits will choke the business.

That capex choice is management explicitly betting they can keep shipping at scale out of China, not signaling an existential block.

[Value Chain Koala @valuechainkoala]

The AI silicon war has exactly one company with no losing scenario: $MRVL

Custom ASIC wins: Marvell designed them. $NVDA GPUs win: Marvell XPUs run inside the same NVLink clusters.

Either way, every optical module connecting to any of them runs Marvell's DSP at ~70% market share.

The duopoly with $AVGO in custom silicon design services isn't a competition; both are fully booked across all four hyperscalers simultaneously.

More total AI silicon investment expands both businesses in parallel.

[Value Chain Koala @valuechainkoala]

@Sam_Badawi Pricing SpaceX at $1.5–1.8T on sub‑$20B revenue and negative net income, while $AMZN is $2.6T on a path to $1T revenue by 2028, isn’t “visionary”, it’s delusional multiple expansion on hype rather than cash flow.

[Value Chain Koala @valuechainkoala]

@aleabitoreddit When an entire critical industry is in the headlines, this is exactly what you get: tiny sentiment shocks trigger outsized panic and cascade selling, even when the fundamental bottlenecks and demand story haven’t changed at all.

[Value Chain Koala @valuechainkoala]

375-layer NAND headlines $TEL, but the real tell is $LRCX .
You can’t build 375 layers without an etch tool drilling holes cleanly through every one of them, and Lam owns that ultra‑deep 3D NAND etch step with no realistic substitute, so every new TEL deposition win at these heights is implicitly a structural win, and widening moat, for $LRCX.

[Value Chain Koala @valuechainkoala]

SK is basically signaling that the bottleneck has moved to the equipment layer, and she’s paying up to lock capacity no matter what.

If SK is raising procurement prices for key tools while $BESI is talking about 3 memory players qualifying hybrid bonding for HBM4 and stretching lead times from ~15 weeks to six months, that’s the same message twice: equipment is the choke point, and the leaders are pre‑buying the bottleneck.

[Value Chain Koala @valuechainkoala]

$LITE CEO already said CPO capacity is sold out through 2028, and $AXTI CEO has been clear demand for InP is running ahead of supply, so tighter timelines and slippage were always the base case, not a shock.

Calling it “they won’t make it in time” now feels less like new information and more like an institutional player talking their book to shake out weak hands and reload at better prices.

[Value Chain Koala @valuechainkoala]

$LITE CEO already said CPO capacity is sold out through 2028, and $AXTI CEO has been clear demand for InP is running ahead of supply, so tighter timelines and slippage were always the base case, not a shock.

Calling it “they won’t make it in time” now feels less like new information and more like an institutional player talking their book to shake out weak hands and reload at better prices.

[Value Chain Koala @valuechainkoala]

Samsung has no reason to fix the conventional DRAM shortage. Here's why.

At +90-95% DRAM prices, Samsung's semiconductor business runs at peak margins. Samsung's mobile business absorbs the higher memory cost, but at the corporate level, the two roughly offset.

Adding supply would compress the pricing that makes the semiconductor unit profitable. Withholding it keeps the margin premium intact.

The world's largest DRAM producer is also the world's largest DRAM buyer. That's not a supply relief valve. That's a structural ceiling. Other producers benefit from Samsung's rational inaction.

[Value Chain Koala @valuechainkoala]

Most people are grading Samsung on today’s CoWoS race, but it’s quietly setting up for the next one.

As AI accelerators grow, 300mm wafers run out of room; Samsung’s shift to large rectangular panels (PLP) is a bet on the moment when only panel‑level packaging can physically fit the biggest next‑gen AI chips.

[Value Chain Koala @valuechainkoala]

@aleabitoreddit $SIVE now sits inside $YSS as a beamforming IC supplier, turning LEO/defense into a structural annuity on top of its AI datacenter laser demand.
two secular ramps funneling through one InP bottleneck.

[Value Chain Koala @valuechainkoala]

Innolight is the 1.6T leader today, but it’s a China‑origin box assembler buying EMLs from $LITE with max geo risk.
$JBL is basically that wallet being rerouted into a US‑origin stack, inheriting Intel’s qualified SiPh lines, while still trading at contract‑manufacturer multiples with $SIVE’s H1 2027 lead times as your tell that the pluggable ramp is real.

[Value Chain Koala @valuechainkoala]

DRAM contracts +90-95% QoQ.
Who captures that pricing without spending billions converting to HBM?

Nanya Technology.
Pure-play conventional DDR4/DDR5. No HBM exposure. No TSV CapEx.

Just a DRAM producer in a market that repriced 90-95% because its three largest competitors allocated capacity to AI instead.

$SNDK invested $1B in Nanya in 2026. That position is now sitting inside one of the most favorable conventional DRAM pricing environments in history.

[Value Chain Koala @valuechainkoala]

@tengyanAI $BE 800V DC fuel cells power data centers without waiting 140+ weeks for grid interconnect hardware, so they don’t face the same delivery pressure and effectively sell time, not just electrons.

[Value Chain Koala @valuechainkoala]

@StockSavvyShay $BE is the energy backbone behind $NBIS data centers, so every new $NVDA ‑powered campus they announce is effectively incremental, contracted demand for $BE 800V DC fuel‑cell architecture.

[Value Chain Koala @valuechainkoala]

@aleabitoreddit $BE data center fuel cells sit on a native 800V DC architecture, which gives them a real moat as hyperscale campuses shift from legacy AC to high‑voltage DC backbones built for AI‑era power density.

[Value Chain Koala @valuechainkoala]

Conventional DRAM prices just became one of the best real-time AI capex indicators available.

GPU production pulls HBM capacity from shared DRAM fabs.

Less capacity for conventional → prices spike.

The conventional DRAM price sits downstream of the HBM allocation decision, which sits downstream of GPU production rates.

When it reverses, it won't read as "consumer electronics softness." It'll be an early signal that $NVDA or $AMD shipment velocity is slowing weeks before sell-side picks it up.

+90-95% QoQ says AI production is still running full.