The Photonics Ledger: Sivers, CPO, and the Unpriced Infrastructure of the AI Trade
Data indicates a supply bottleneck in a layer of the technology stack that most AI narratives ignore. Sivers Photonics (SIVE), a London AIM-listed specialty foundry for III-V compound semiconductors and silicon photonics, is running capacity utilization above 90% while average selling prices climb. The company's order book carries economic value that the market has not fully priced. This is not a crypto story in the conventional sense. But it is a ledger story. The physical infrastructure that settles the AI compute trade is the same infrastructure that will settle the tokenized AI economy. When I audited AI-agent trading protocols in 2026, I found that the latency arbitrage problem was not a software problem. It was a hardware problem. The photonics layer determines who sees the data first.
Sivers operates in the specialty foundry segment. Its core competency is InP (indium phosphide) active device integration combined with silicon photonics. The company supports hybrid integration - bonding InP gain media onto silicon photonic passive components. This is the technical path for Co-Packaged Optics (CPO), the packaging architecture that AI data centers require to reduce power consumption and increase bandwidth density.
The market context is favorable. CPO penetration is projected to grow from under 5% in 2024 to 20-30% by 2028. Ayar Labs, the Intel-backed CPO pioneer, has expansion plans through 2028. Sivers supplies optical engine chips to CPO manufacturers. The company also produces External Laser Sources (ELS) in partnership with O-Net, a Chinese optical module maker.
The supply chain position is high-value. Optical chips account for 30-50% of the BOM cost in optical modules. The profit pool distribution in optical communications: optical chips and devices capture 35-40%, optical modules 30-35%, equipment integration 20-25%, materials 5-10%. Sivers sits in the richest segment.
The company's upstream dependencies are concentrated. InP substrates are dominated by Japanese suppliers - Sumitomo Electric, JX Nippon Mining. High-precision lithography equipment comes from ASML, Canon, and Nikon. Epitaxial growth equipment from Aixtron and Veeco. The supply chain is globalized but concentrated. The company's downstream customer concentration is equally high - the top five customers likely account for 70-80% of revenue, with Ayar Labs or O-Net as the largest. The mention of six new pluggable customers suggests diversification is underway, but the base remains narrow.
The technology assessment requires precision. In silicon photonics foundry capability, Sivers trails TSMC's COUPE platform by approximately 2-3 years. TSMC's COUPE (Compact Universal Photonic Engine) is expected to enter mass production in 2025. GlobalFoundries' 45nm silicon photonics platform is also ahead. However, in InP active integration, Sivers is in the industry's first tier, competing with Intel and Broadcom's silicon photonics teams.
This distinction matters. The InP active integration capability is the bottleneck in CPO manufacturing. Laser performance consistency and co-design with CPO packaging processes are the core barriers. Sivers' know-how in InP epitaxial growth is its moat. The company's yield in InP active integration is estimated at 70-85%, which is lower than mature silicon photonics platforms (85-95%) but competitive for the segment. Yield is the lever that determines unit economics. In a supply-constrained market, every percentage point of yield improvement translates directly into margin expansion.
The financial picture is more complex. The company's gross margin is estimated at 30-40%, which is healthy for a specialty foundry. IQE, the III-V foundry leader, runs at 25-30%. The supply bottleneck and rising ASPs suggest margin expansion. However, the company's cash flow is tight. Capital expenditure requirements for a potential US fab would pressure the balance sheet. The estimated ROIC of 3-5% is below the WACC of 10-12%. The company is not currently creating value. This is the structural tension: the technology is right, the market timing is right, but the balance sheet is thin.
The Serenity critique is the central tension. Serenity, an activist investor, argues that Sivers' focus on the Swedish market constrains its valuation. The logic: Swedish market investor structure - retail and local institutions - cannot properly price CPO growth stories. US investors, dominated by institutions, would assign a growth multiple. The recommendation is to pivot market focus to the US.
This critique has merit. The valuation metrics support it. Sivers trades at 5-8x price-to-sales, which is a premium to IQE's roughly 3x. But Serenity argues the order book's economic value is not reflected. The supply bottleneck means orders carry higher margins. The six new pluggable customers suggest customer diversification is underway. The market is pricing the company as a Swedish small-cap, not as a CPO supply chain play.
The competitive threat is real. TSMC's COUPE platform, with its scale, capital, and customer relationships, could reshape the CPO optical engine market. The probability of TSMC successfully mass-producing COUPE in 2025 is 60-70%. This is the primary risk to Sivers' thesis. TSMC does not need to win every CPO design. It only needs to win the top three hyperscaler accounts to compress Sivers' addressable market significantly.
The two-fab question adds another layer. The mention of two wafer fab capacity allocation suggests Sivers may have or be planning a dual-site footprint - Sweden and the US. The US site would be strategically closer to CPO customers like Ayar Labs. It would also qualify for CHIPS Act support. But a new fab means 12-18 months from equipment installation to mass production, and depreciation would pressure gross margins by 3-5 percentage points until the fab reaches 60-70% utilization. The capacity allocation decision is a strategic choice: prioritize high-value CPO customers or diversify across pluggable module makers. In a supply-constrained environment, allocation is strategy.
The inventory cycle is favorable. The optical communications industry runs on 2-3 year inventory cycles: 2019 destocking, 2021 restocking, 2023 destocking, 2024 restocking. The current cycle is in the restocking phase, driven by AI demand. Supply bottlenecks are expected to persist until 2025-2026, when new capacity comes online. The AI demand driver is structural, not cyclical. AI training clusters like NVIDIA GB200 NVL72 require high-speed optical interconnects, with per-rack optical module value rising from thousands of dollars to tens of thousands. This is not a marginal increase. This is a step function.
The geopolitical layer is manageable but present. Sivers is not on any US entity list. Optical communications chip manufacturing is not subject to advanced process export controls - no EUV, no sub-3nm. However, China's export controls on gallium and germanium do not directly affect Sivers, whose core material is InP. The indirect competitive pressure comes from China's Big Fund Phase III, which targets advanced packaging and optical communications chips, potentially nurturing domestic competitors like Yuanjie Technology and Accelink. The decoupling risk is rated low-to-moderate, but the indirect competitive pressure is real.
The US pivot narrative is overrated. Moving market focus to the US does not change the fundamental competitive dynamics. The real issue is not where the stock is listed or where the investor base sits. The real issue is whether Sivers can maintain its InP integration advantage as TSMC scales COUPE.
There is a counter-intuitive angle: Sivers' position as a neutral foundry - a UK/Swedish company - has strategic value in the US-China technology decoupling. The company can serve both US customers like Ayar Labs and Chinese customers like O-Net without triggering compliance concerns. This neutrality is an asset that the pivot-to-US narrative would erode. The market is focused on the narrative of where Sivers should be listed. The actual value is in the supply chain position - the InP epitaxial know-how, the CPO customer bindings, the capacity scarcity. We mapped the water, not the wave.
The deeper risk is not valuation. It is the assumption that CPO demand will materialize on schedule. Ayar Labs' 2028 expansion is a plan, not a contract. If AI capex cycles slow, the CPO inflection point shifts right. The supply bottleneck that currently gives Sivers pricing power could reverse quickly. The company's high customer concentration amplifies this risk. A single design win loss at a major CPO integrator would remove 20-30% of revenue.
A ledger is a confession written in code. Sivers' ledger shows a company with real technical assets, real order momentum, and real competitive threats. The CPO inflection point is real. The question is whether Sivers can execute before TSMC's COUPE platform reshapes the market. The market narrative about Sweden versus the US is a distraction. The technical execution is the variable that matters. I have seen this pattern before - in the 2022 Terra collapse, the market focused on narratives while the math was already broken. Here, the math is sound but the competition is fierce. Watch the InP yield data, not the stock exchange listing. The structural integrity of the technology will determine the speculative value of the equity.