Tower Semiconductor Jumped 9.9% on a Deal Signed Two Weeks Ago — and Still Trades 42% Below Its High

The binding constraint on an AI cluster is no longer only how many accelerators you can buy. It is how fast you can move data between them, and that traffic is converted by analog and photonic silicon: the silicon germanium drivers and transimpedance amplifiers inside the rack, and the photonic integrated circuits that modulate light across the hall. That layer is sourced from mature and specialty nodes, not leading-edge fabs, and it is now the part of the build-out that is genuinely rationed.

Tower's revenue mix shows how far the shift has gone. In the second quarter of 2026, RF infrastructure reached 49% of corporate revenue with silicon photonics alone running at a $680 million annualised rate against $180 million a year earlier, while power management, where management pointed to demand for power delivery in high-performance computing, added another 14%. Roughly two-thirds of revenue now rides on AI data-centre content, against a company that was a diversified analog foundry two years ago.

On October 13, Tower and Semtech expanded a partnership that had already run for twenty years. Under the new multi-year agreement Semtech invests in securing dedicated silicon germanium capacity at Tower's facilities in the United States, Israel and Japan, explicitly for AI data-centre optical and copper interconnect, from 800G and 1.6T today through 3.2T and beyond. What the agreement proves is that dedicated 200mm silicon germanium capacity has to be reserved well ahead of demand, and that a fabless supplier is willing to fund that reservation.

The deal also says something the headline hides. The reserved capacity spans three fabs in three countries, and not one of them is a leading-edge node. Silicon germanium drivers for 800G and 1.6T links are 200mm specialty work, while the industry's 300mm photonics build-out is happening in Japan and Italy. Tower's answer to a rationed leading edge is breadth of geography and node, not depth of node.

Here is the part of the story that is easy to miss. Tower's trailing twelve-month operating cash flow was $873.6 million against net income of $289.6 million, and $232.0 million of that gap came from working capital, which is the silicon photonics customers paying in advance to hold capacity. Deferred revenue on the balance sheet went from nothing at the end of 2025 to $321.4 million by June 30, 2026. Strip out the prepayments and trailing free cash flow falls from $304.6 million to roughly $73 million, which is the real earnings power of the business while the build-out runs.

Two facts reframe the geopolitical discount. First, Israel's contribution is one 200mm fab: Fab 1's legacy 150mm flows were discontinued in the first quarter of 2025, leaving Fab 2 as the only Israeli wafer facility. Second, every 300mm photonics wafer in Tower's plan is being built in Japan, at the Arai conversion, the expanded Fab 7 and a new adjacent 300mm building, with Italy's Agrate sharing the 300mm base. The wafer capacity Tower is adding sits in Japan, the United States and Italy, not in Israel, which narrows rather than removes the risk premium.

Management's 2028 model, $3.6 billion of revenue and $1.2 billion of net profit, is arithmetic rather than aspiration, but it is demanding arithmetic. Getting from $1.57 billion of2025 revenue to $3.6 billion requires a 32% revenue compound rate for three consecutive years, and the first step is already largely locked: the $520 million third-quarter guide annualises to $2.08 billion. Gross margin has to travel from 26.9% to 45% and operating margin from 16.2% to 38%, which assumes 85% utilisation across the fabs, operating expenses falling to about 7% of revenue and R&D spending rising 40%.

At $184.14 the market pays $20.9 billion for the company, or 12.2 times trailing revenue and 72 times trailing earnings, for a business currently earning a 7.4% return on invested capital with capex at a third of revenue. The average target across covering analysts is $312.93, roughly 70% above the close. The shares sit 42% below their 52-week high, so the market has already taken a substantial discount off this story once, and the question is no longer whether the narrative is exciting.

In days and quarters, everything narrows to the November 2 report. The company guides to $520 million of revenue plus or minus 5% against a consensus $1.07 of earnings per share, and the second quarter delivered $460.1 million and $0.80. The number that actually matters is the silicon photonics run rate, which management targets above $1 billion by the fourth quarter, with the capacity behind it running at more than three times second-quarter shipment levels by late 2026.

Over one to three years, the bet is that 300mm photonics becomes a Japanese and Italian business run by an Israeli company. The Arai fab is due to be production-ready for 300mm silicon photonics and advanced packaging in the fourth quarter of 2027, and a second Japanese building is targeted for the fourth quarter of 2028, intended to multiply Japan's 300mm output fourfold. Everything in the 2028 model sits on that Japanese build, and none of it exists yet.

The upstream constraint is epitaxy and government money. In June, IQE signed a multi-year agreement under which Tower commits to minimum purchases of indium phosphide epitaxial wafers, with reciprocal volume commitments from IQE, for platforms running 200Gb/s and future 400Gb/s per lane, and Japan's government is underwriting roughly $1 billion of the Japanese build. Tower's ramp is underwritten by its own suppliers and its host government before the wafer starts, which is a materially different risk profile from a foundry funding capacity out of its own cash flow.

Downstream, the names are more visible. NVIDIA has been collaborating with Tower on 1.6T data-centre optical modules since February 2026, and Tower's silicon photonics enables up to double the data rate of prior silicon photonics solutions. Zhongji Innolight expanded its collaboration with Tower and has been ramping next-generation silicon photonics for AI and data centres. Marvell and Tower have shipped more than five million coherent photonic integrated circuits, Coherent partnered on the PH18 platform, and private challenger NewPhotonics began high-volume shipments of laser-integrated optical engine PICs on Tower's PH18DA platform in September, with 6.4T chipsets planned for volume in the first half of 2027. The module makers and hyperscalers are pulling Tower's silicon onto their roadmaps while Tower's customers fund the expansion.

The synthesis is straightforward. The scarce commodity in the AI build-out has shifted from compute silicon to the analog and photonic content that moves data between accelerators, and Tower has pre-sold a meaningful slice of that content for 2027 while financing the capacity behind it with customer money. The investable question is not whether Tower is an AI company, but whether a 45% gross margin arrives on schedule, and the next five weeks of evidence start on November 2.