The order book is real. The market is squinting at the wrong fab.
Supply bottlenecks don't lie. Narratives do. Sivers Photonics (SIVE) sits inside a genuine capacity squeeze: utilization high enough to make this a seller's market, average selling prices climbing, six new pluggable customers through the door, an external laser source (ELS) product running with O-Net, and a CPO customer — Ayar Labs — whose 2028 expansion is the industry's most public timestamp for co-packaged optics going mainstream. The market's response: price it like a Swedish specialty foundry.
I've seen this pattern before. In 2017, I spent four months auditing the Golem ICO distribution contract, parsing assembly opcodes until I found an integer overflow in the batch claim function. The patch shipped before mainnet. The lesson stuck: trust must be cryptographically enforced, not socially promised. Equities obey the same rule. SIVE is a claim on a supply chain, not a narrative. But it's priced like the narrative.
Context: What Sivers Actually Is
Sivers is not a CMOS logic foundry. No 5nm, no 3nm, no EUV. It is a specialty III-V and silicon photonics wafer fab. The relevant metrics are waveguide dimensions, coupling efficiency, and integration density. Its core capability is hybrid integration — mounting InP gain media on silicon photonic passives. That is the technical backbone for both pluggable optical modules and co-packaged optics (CPO), where the optical engine sits on the same substrate as the switch ASIC.
Positioning matters. TSMC's COUPE platform targets 2025 mass production. GlobalFoundries fields a 45nm silicon photonics platform. Against that baseline, Sivers trails by 2-3 years on large-scale CMOS-compatible silicon photonics — roughly one to two generations on on-chip device count. But in InP active integration — the laser and amplifier layer — it sits in the first tier, competing against Intel's and Broadcom's internal teams.
The 2022 LUNA collapse taught me to model confidence thresholds. The UST death spiral was inevitable once confidence dropped below 60%. The seigniorage model didn't fail on code; it failed on a social promise. I test every supply chain the same way. The AI optical trade is a confidence game on a longer clock. The promise: hyperscaler capex stays elevated until CPO scales.
Core: Where the Analysis Bites
Start with yield. Silicon photonics foundry yield runs 85-95%; InP active integration runs lower, roughly 70-85%. Yield is the hidden P&L. In a supply bottleneck, order counts mean less than ASP times yield per wafer. A 10-point yield swing on InP integration moves unit economics more than a press release about a new customer. Sivers does not publish yield data. That silence is the first red flag — and the first clue. Tracing the gas leaks before the code compiles. The wafers play nice in the demo. The question is whether they survive the volume ramp.
The value chain position is the real asset. Optical chips run 30-50% of module BOM cost and sit in the fattest profit layer in optical communications — roughly 35-40% of the pool, versus 30-35% for modules and 20-25% for integration. Sivers occupies the richest layer. But the dependence is heavy: InP substrates are dominated by Japanese suppliers; MOCVD and MBE epitaxy tools come from Aixtron and Veeco; there is no mature substitute. This equipment is not export-controlled — no EUV, no sub-3nm trigger — which keeps the chain functional but dependent. Meanwhile, China's state-backed funds are pushing localization in exactly this layer, funding domestic InP and silicon photonics competitors.
Geopolitics frames both the upside and the failure mode. Sivers is not on any US entity list, and its toolset sits outside the advanced-node export-control blast radius. That neutrality is an asset: a UK/Sweden base can serve Western CPO champions and Chinese module makers without triggering compliance reviews. The same neutrality is a liability if the US expands AI-infrastructure controls into co-packaged optics. The InP substrate chain, concentrated in Japan, is the physical choke point. It is not a strategic military item, so the risk stays theoretical — but theoretical risks have a habit of arriving at the worst possible moment.
Capacity allocation is strategy. The "two wafer fabs" language hides the real question: which customers get the wafers? CPO customers or pluggable module customers? When capacity is the constraint, allocation is the strategy. A US fab is not just a geopolitical hedge — it is the key to US-based customers and, potentially, CHIPS Act funding. But a new fab needs 12-18 months from tool-in to revenue. Depreciation will shave 3-5 percentage points off gross margin. Breakeven sits at 60-70% utilization. With capex intensity at 20-30% of revenue, free cash flow stays negative, and dilution becomes a real probability. The market is not pricing dilution. It is pricing the release cycle.
Demand is the strongest leg. AI racks have lifted optical interconnect value from thousands to tens of thousands of dollars per cabinet. LightCounting puts CPO in the tens of billions by 2028. Penetration moves from under 5% in 2024 to 20-30% by 2028. Supply bottleneck plus rising ASP equals pricing power. That is the core of the "economic value of orders" argument — correct, with a timing caveat. Ayar's 2028 expansion is the timestamp. Everything before that is pipeline, not revenue.
The competitive map is the tension. IQE leads III-V foundry; TSMC leads silicon photonics. Sivers holds maybe 10-15% of InP active integration but only 3-5% of silicon photonics. R&D intensity at 15-20% of revenue signals focus — but the absolute dollars are millions against billions. That is the small-foundry paradox: Sivers outspends proportionally but cannot outspend absolutely. Its moat is not scale. Its moat is know-how in InP epitaxy and binding relationships with CPO partners. The model didn't price what happens if that know-how becomes a commodity — or if a top-five customer walks.
Financials split it clean. Gross margin in the 30-40% range is healthy for a specialty foundry. But the denominator matters: ROIC of 3-5% against a WACC of 10-12% destroys value on current math. The valuation — PS at 5-8x, EV/EBITDA at 20-30x — is pricing 2028 today. Operating cash flow will likely run below net income, and free cash flow will be negative through the capex cycle. The bull case requires the bottleneck to persist longer than the capex cycle. That is a narrow window.
Contrarian: The Easy Narrative Is the Wrong One
The popular read says pivot to the US and the multiple re-rates. That is a narrative fix, not a structural one. US institutions give growth multiples; Swedish retail does not. The valuation gap is real — but the arbitrage is on multiples, not on fundamentals. Customers do not care where the listing lives. They care about InP consistency and delivery. Moving the center of gravity changes the investor base, not the wafer starts.
TSMC's COUPE entry is the other comfortable story: the death knell. Wrong again. TSMC entering CPO validates the market. The genuine threat is concentration. Six new customers sound like diversification; they mask a top-five concentration around 70-80% of revenue. Ayar Labs is both the bull case and the single point of failure. If Ayar stumbles, the order book compresses in a quarter.
Then there is the subsidy layer. This resembles DeFi liquidity mining on a longer clock: subsidized TVL evaporates when the incentives stop. AI capex is the incentive. Stop the capex — or let confidence slip — and orders revert to goodwill with a delivery date. Liquidity is just patience with a time limit. The market's patience is set to Ayar's 2028 clock.
Silence between the blocks tells the real story. The missing disclosures — yield, customer names, fab funding, utilization — are the actual data points. The published narrative is the noise. Two weeks in the lab, one second in the field. The technical case is real. The financial base is thin. The valuation sits exactly between the bottleneck and the narrative.
Takeaway
Watch three things: named CPO design wins, Ayar's 2028 expansion converting to purchase orders, and whether the US fab shows up on a balance sheet instead of a slide deck. If the bottleneck holds, the PS multiple re-rates. If COUPE ships on schedule and a top-five customer walks, the discount becomes a chasm. The InP layer is the moat. The customer list and the capex math are the failure points. The signal to trade is not the next press release. It is the wafer starts.