Israeli specialty foundry manufacturing analog, mixed-signal, silicon photonics, and power management chips. Leopold holds $85M and has been building. NOT a leading-edge foundry -- they make the chips that make AI servers actually work. | Analysis date: 2026-03-12
Leopold Aschenbrenner has been building a position in Tower Semiconductor: 470,600 shares bought in Q3 2025 at ~$72/share, expanded to 723,004 shares in Q4 2025 at ~$117/share. Average cost basis ~$80. Current value ~$85M, ~2.1% of his portfolio. This is a pure infrastructure play on the AI supply chain -- not the glamorous GPU layer, but the physical layer beneath it. Every AI server needs power management ICs, every data center interconnect needs silicon photonics transceivers, and Tower is one of the very few foundries in the world that can manufacture these specialty chips at scale. Intel tried to acquire Tower for $5.4B in 2022 and failed. The stock has gone from $28 to $149 in 12 months. The question: what is the bull case market cap, what is the 10x entry price, and are we near it?
Tower Semiconductor is not competing with TSMC, Samsung, or Intel on leading-edge logic chips (3nm, 5nm). They operate at mature nodes (130nm to 65nm, some 45nm) making specialty analog and mixed-signal chips that digital foundries cannot and do not want to make. These chips do not shrink with Moore's Law -- they get more critical as digital chips get more powerful.
| Silicon Photonics (SiPho) | Optical transceivers for data center interconnects. Converting electrical signals to light and back. Critical for 400G/800G/1.6T links between AI servers. |
| Power Management ICs (PMICs) | BCD (Bipolar-CMOS-DMOS) processes for voltage regulators, power converters. Every GPU needs multiple PMICs to deliver clean power. |
| RF CMOS / SiGe BiCMOS | Radio frequency chips for 5G, WiFi, cellular infrastructure. High-frequency analog circuits that digital foundries cannot make. |
| CMOS Image Sensors | Camera sensors for smartphones, automotive, industrial vision. Specialty pixel-level processing. |
| MEMS / Non-Imaging Sensors | Accelerometers, gyroscopes, pressure sensors. Physical world interfaces. |
| High-Voltage CMOS | Up to 700V ultra-high voltage. Industrial, automotive, power grid applications. |
An NVIDIA H100 GPU is useless without:
The AI data center is a system. The GPU gets all the attention, but for every $30,000 GPU, there are hundreds of dollars of specialty analog chips that enable it to function. Tower makes those chips. As AI GPU deployments scale 10x-100x, the demand for these analog companion chips scales with it -- often faster, because power delivery and interconnect complexity grow super-linearly with GPU count.
Tower's silicon photonics platform (PH18) is particularly important for the AI thesis. From the 20-F:
"We currently have a high-volume, production 200mm platform (PH18) that has benefited from AI-driven growth in optical interconnects, making it a leading SiPho platform in the 400Gb/s to 800Gb/s market. We have begun production of next-generation products targeting 1.6Tb/s and are developing and prototyping various advanced technologies aiming at 3.2Tb/s, as well as co-packaged optics (CPO) and integrated laser applications for the future."
Translation: Tower is already in production for 800G optical transceivers used in AI data centers, ramping 1.6T, and developing 3.2T. These are the interconnects that link thousands of GPUs together in AI training clusters. The market for these is exploding. Tower is one of only a handful of foundries worldwide that can manufacture silicon photonics at production scale. This is the single most directly AI-exposed part of their business.
Tower operates a geographically diversified network of fabrication facilities. This is unusual for a specialty foundry and provides both capacity flexibility and geopolitical risk mitigation.
| Fab | Location | Wafer Size | Key Technologies | Status |
|---|---|---|---|---|
| Fab 1 | Migdal Haemek, Israel | 150mm | Legacy processes | Discontinued Q1 2025 |
| Fab 2 | Migdal Haemek, Israel | 200mm | Analog, RF SOI, power, mixed-signal, SiPho (PH18) | Active -- primary fab |
| Fab 3 | Newport Beach, CA | 200mm (0.50-0.13um) | SiGe, SiPho, analog, RF, high-voltage, SOI | Active -- lease expires 2027 |
| Fab 7 | Uozu, Toyama, Japan | 200mm | Specialty processes (TPSCo) | Active |
| Fab 9 | San Antonio, TX | 200mm | Specialty (acquired from Maxim 2016) | Active |
| Fab 10 | Agrate, Italy (shared w/ ST) | 300mm | Advanced specialty -- 1/3 of ST cleanroom | Volume production started Q4 2024 |
| Fab 11 | New Mexico, USA (Intel facility) | 300mm | Tower processes on Intel capacity corridor | Equipment procurement phase -- not yet qualified |
| Tonami CD | Tonami, Toyama, Japan | 200mm | TPSCo operations | Active |
The 300mm transition is the strategic play. Moving from 200mm to 300mm wafers reduces cost per die by 30-40% while increasing capacity. Fab 10 (Italy) and Fab 11 (Intel) together represent a massive increase in Tower's competitive position -- 300mm specialty fabs are extremely rare and take years to build.
| February 15, 2022 | Intel and Tower announce definitive agreement. Intel to acquire all Tower shares for $53/share cash (~$5.4B). Intel's goal: use Tower's specialty foundry expertise to build out its Intel Foundry Services (IFS) business with analog/mixed-signal capabilities. |
| 2022-2023 | Deal requires Chinese regulatory (SAMR) approval. China does not approve within the required timeframe. The deal is caught in US-China tech cold war -- China likely blocked it as retaliation for US semiconductor export controls. |
| August 16, 2023 | Intel and Tower announce termination of the merger agreement. Closing conditions not met because "certain regulatory approvals were not received." |
| Termination Fee | Intel pays Tower a $313.5 million termination fee (net of associated costs). This one-time payment inflated FY2023 operating profit to $547M and net income to $518M. Stripping it out, normalized FY2023 operating profit was ~$234M. |
| September 2023 | Instead of an acquisition, Tower and Intel sign a capacity corridor agreement -- Tower will install its own equipment at Intel's New Mexico 300mm fab (Fab 11). Tower invests up to $300M. This achieves part of what the acquisition would have: Tower gets 300mm capacity, Intel gets a customer for its underutilized fabs. |
| Metric | FY2020 | FY2021 | FY2022 | FY2023 | FY2024 | TTM (Q3 2025) |
|---|---|---|---|---|---|---|
| Revenue | $1,266M | $1,508M | $1,678M | $1,423M | $1,436M | $1,566M |
| YoY Growth | -- | +19% | +11% | -15% | +1% | +14% |
| Cost of Revenue | $1,032M | $1,179M | $1,211M | $1,069M | $1,097M | -- |
| Gross Profit | $233M | $329M | $466M | $354M | $339M | -- |
| Gross Margin | 18.4% | 21.8% | 27.8% | 24.8% | 23.6% | 23.2% |
| R&D Expense | $78M | $85M | $84M | $80M | $79M | -- |
| SG&A Expense | -- | $77M | $80M | $72M | $75M | -- |
| Operating Income | $91M | $167M | $312M | $547M* | $191M | -- |
| Operating Margin | 7.2% | 11.0% | 18.6% | 38.5%* | 13.3% | 16.1% |
| Net Income | $82M | $150M | $265M | $518M* | $208M | -- |
| EPS (Diluted) | $0.76 | $1.36 | $2.39 | $4.66* | $1.85 | $1.94 |
| EBITDA | -- | $433M | $590M | $847M* | $488M | -- |
*FY2023 includes $313.5M Intel merger termination fee. Normalized operating income was ~$234M and normalized net income was ~$205M.
| Quarter | Revenue | Gross Profit | Operating Income | Net Income | Diluted EPS |
|---|---|---|---|---|---|
| Q4 2024 | $387M | $87M | $46M | $55M | $0.49 |
| Q1 2025 | $358M | $73M | $33M | $40M | $0.35 |
| Q2 2025 | $372M | $80M | $40M | $47M | $0.41 |
| Q3 2025 | $396M | $93M | $51M | $54M | $0.47 |
| Q4 2025 (est.) | ~$410-420M | -- | -- | -- | $0.70* |
*Q4 2025 diluted EPS from yfinance. Revenue trajectory shows clear acceleration: each quarter higher than the last.
After the post-cycle trough in FY2023-24 (semiconductor downturn), Tower's revenue is accelerating quarter-over-quarter: $358M, $372M, $396M, and likely $410-420M in Q4 2025. This puts the annualized run rate at $1.55-1.68B, approaching the FY2022 peak of $1.68B. If this trajectory holds, FY2026 revenue could reach $1.7-1.9B, setting a new all-time high.
The growth drivers: (1) AI-driven silicon photonics demand, (2) automotive and industrial recovery, (3) Fab 10 (Italy) ramping volume production, (4) power management IC demand from data center buildout. All four end markets are expanding simultaneously.
| Item | FY2024 | FY2023 | FY2022 | FY2021 |
|---|---|---|---|---|
| ASSETS | ||||
| Cash & Equivalents | $272M | $261M | $341M | $211M |
| Short-Term Investments | $946M | $976M | $665M | $554M |
| Cash + ST Investments | $1,218M | $1,236M | $1,006M | $765M |
| Accounts Receivable | $212M | $154M | $153M | $142M |
| Inventory | $268M | $283M | $302M | $235M |
| Total Current Assets | $1,760M | $1,709M | $1,495M | $1,196M |
| PP&E (Net) | $1,295M | $1,166M | $973M | $891M |
| Goodwill | $7M | $7M | $7M | $7M |
| Intangibles (ex goodwill) | $6M | $8M | $10M | $15M |
| Total Assets | $3,080M | $2,919M | $2,548M | $2,231M |
| LIABILITIES | ||||
| Current Liabilities | $285M | $277M | $387M | $276M |
| Long-Term Debt | $132M | $173M | $210M | $231M |
| Total Debt | $181M | $232M | $272M | $315M |
| Total Liabilities | $440M | $491M | $659M | $616M |
| EQUITY | ||||
| Stockholders' Equity | $2,653M | $2,432M | $1,891M | $1,622M |
| Tangible Book Value | $2,641M | $2,418M | $1,874M | $1,600M |
This is the opposite of CORZ. No bankruptcy, no negative equity, no convertible debt, no warrant liabilities. Tower's balance sheet alone is worth roughly $23.70/share in tangible book value. The floor on this company is high.
| Cash Flow | FY2024 | FY2023 | FY2022 | FY2021 |
|---|---|---|---|---|
| Operating Cash Flow | $449M | $677M* | $530M | $421M |
| Capital Expenditures | -$436M | -$445M | -$366M | -$314M |
| Free Cash Flow | $13M | $232M* | $163M | $107M |
| Depreciation | $266M | $258M | $293M | $271M |
| Net Debt Repayment | -$32M | -$32M | -$78M | -$77M |
*FY2023 OCF includes ~$313M Intel termination fee. Normalized OCF was ~$360M.
Tower is in a heavy capex phase ($436M-445M/yr) as it expands Fab 10, equips Fab 11, and migrates technologies from Fab 3. This is eating most of the operating cash flow, leaving minimal free cash flow in FY2024. But several things are important:
| Company | Revenue | Market Cap | Focus | Nodes | Tower Comparison |
|---|---|---|---|---|---|
| Tower (TSEM) | $1.6B | $13.3B | Pure specialty: SiPho, PMICs, RF, image sensors | 130nm-45nm | -- |
| GlobalFoundries (GFS) | $7.4B | $22B | Specialty + mainstream (22nm-12nm) | 350nm-12nm | GFS 4.5x revenue but only 1.7x market cap. Tower trades at premium multiple. |
| TSMC (TSM) | $90B | $850B | Leading edge + some specialty | 350nm-3nm | Different league. TSMC does everything but the specialty market is a small fraction of their business. |
| X-Fab (XFAB.PA) | $0.8B | $2.5B | Analog, MEMS, automotive | 180nm-130nm | Smaller, less diversified than Tower. Direct specialty competitor. |
| Hua Hong (1347.HK) | $1.6B | $5B | Power discrete, specialty CMOS | 350nm-55nm | Chinese competitor. Geopolitical risk for customers needing non-China supply. |
| Vanguard (5347.TW) | $1.1B | $5B | Mixed-signal, power, display drivers | 350nm-90nm | TSMC subsidiary. Less technology breadth than Tower. |
GlobalFoundries is the most direct comparable, but the comparison reveals why Tower may deserve a premium:
| Technology breadth | Tower has deeper specialty capabilities (SiPho, SiGe BiCMOS) than GFS. GFS has broader mainstream capabilities (22nm, 12nm FinFET-like). |
| Silicon Photonics | Both have SiPho platforms. Tower's PH18 is in high-volume 800G production. GFS's fotonix platform competes. This is the key AI battleground between them. |
| Revenue concentration | GFS depends heavily on a few large customers (AMD ~25% of revenue historically). Tower has no single customer above ~15% -- better diversification. |
| Margins | Tower: 23-28% gross margin. GFS: ~24-26% gross margin. Similar, but Tower's margin should expand as 300mm fabs (Fab 10, Fab 11) ramp -- these have structurally lower costs per wafer. |
| Valuation | Tower: 8.5x EV/Revenue, 27x EV/EBITDA. GFS: 3x EV/Revenue, 8x EV/EBITDA. Tower is much more expensive, implying the market prices in faster growth and higher AI exposure. |
| Balance sheet | Tower: net cash $1B, debt-to-equity 6%. GFS: net debt ~$0.7B, more leveraged. Tower wins decisively. |
Summary: Tower trades at a big premium to GFS. This is justified only if the AI/SiPho growth thesis plays out and the 300mm transition succeeds. If Tower's specialty focus proves right (power management + silicon photonics are the bottlenecks in AI infrastructure), the premium is warranted. If the semi cycle softens, Tower's premium could compress toward GFS multiples.
| Entity | Situational Awareness LP |
| Filing Type | 13F (passive -- NOT activist like CORZ) |
| Q3 2025 Entry | 470,600 shares at ~$72.30 |
| Q4 2025 | 723,004 shares at ~$117.42 |
| Shares Added Q4 | +252,404 (+54%) |
| Average Cost Basis | ~$80.18 |
| Current Value (@ $118.54) | ~$85.7M |
| Gain | +47.8% ($26.5M unrealized) |
| Portfolio Weight | ~2.1% |
Leopold's portfolio thesis is clear: AGI requires massive physical infrastructure, and the bottlenecks are in the physical layer. His portfolio includes:
Tower fits the thesis perfectly: Leopold is not buying the GPU companies (NVIDIA) or the cloud companies (Microsoft, Google). He is buying the companies that make the physical infrastructure that GPU companies and cloud companies depend on. Tower makes the silicon photonics chips that connect GPUs together and the power management ICs that deliver electricity to GPUs. Without Tower's chips, NVIDIA's GPUs are expensive paperweights.
He added to the position in Q4 2025 at ~$117. He bought more at prices near today's price, which suggests he thinks there is still significant upside from here.
| Cash + Investments | $1,218M | Massive liquidity buffer |
| Total Debt | $181M | Trivial relative to cash |
| Net Cash | $1,037M | No solvency risk |
| Stockholders' Equity | $2,653M | Strongly positive |
| PP&E | $1,295M | Real, valuable fab assets |
| Goodwill | $7M | Virtually nothing |
| Operating Cash Flow | $449M/yr | Consistently profitable |
| Customers | Diversified | No single customer dominant |
Zero risk: Extremely low. Tower has $1.2B in liquid assets, generates $450M/yr in operating cash flow, has minimal debt, positive equity, real physical assets (semiconductor fabs), a diversified customer base, and operates in an industry with structural demand growth. There is no bankruptcy scenario that makes sense unless the entire semiconductor industry collapses. Even in a severe downturn, Tower's fabs and processes have significant value to acquirers. Intel already valued the company at $5.4B.
Result: This passes the "can't go to zero" test with flying colors.
| Floor Component | Value | Notes |
|---|---|---|
| Tangible Book Value | $2,641M | = $23.68/share. Virtually all real assets. |
| Cash + Investments (liquid) | $1,218M | = $10.92/share in liquid assets alone. |
| PP&E Replacement Value | $1,800-2,500M | Fabs cost far more to build than book value. Gross PP&E is $4.75B. |
| Trough Earnings Power | $1.30-1.80/share | Trough EPS (semi downturn) |
| Trough Multiple | 12-15x | Conservative specialty semi multiple |
| Trough Earnings Floor | $16-27/share | Trough EPS x trough P/E |
| Combined Floor (Asset + Earnings) | $35-55/share | Tangible book + ~50% premium for earnings power |
The floor here is robust. Even in the worst semiconductor downturn, Tower's fabs, cash, and customer relationships are worth at minimum $35-55/share. Intel's $53/share bid in early 2022 -- when the company was less valuable than today -- provides a hard anchor. The floor confidence is HIGH because: (1) minimal goodwill/intangibles, (2) net cash position, (3) diversified customer base, (4) real physical assets that would take 5+ years and $5B+ to replicate.
Working backwards from the bull case: even the mega bull scenario ($60-78B) implies a 10x entry of $6-7.8B (~$54-70/share). At $13.3B, Tower is already well above the 10x entry zone. For comparison:
| Scenario | Revenue | Net Margin | Net Income | P/E | Market Cap |
|---|---|---|---|---|---|
| Current | $1.57B | 14% | $208M | 64x | $13.3B |
| Bull Case 2030 | $4-5B | 25% | $1-1.25B | 30x | $30-38B |
| Mega Bull 2030 | $6-8B | 28% | $1.7-2.2B | 35x | $60-78B |
| What 10x entry requires | $10B+ | 30%+ | $3B+ | 40x+ | $130B+ |
10x requires Tower to grow revenue 6-7x and dramatically expand margins. That would require (a) silicon photonics becoming a $10B+ market with Tower taking 20%+ share, (b) power management ICs for AI scaling massively, (c) Fab 10 and Fab 11 fully ramped at 300mm, (d) margin expansion from 23% gross to 35%+ gross, and (e) the market assigning 30-40x P/E. This is possible over 10 years but low probability.
| Entry Price | Market Cap | Bull Case Target | Multiple | Assessment |
|---|---|---|---|---|
| $30-40/share | $3.4-4.5B | $30-38B | 7-10x | This is the 10x zone. Intel bid $53/share in 2022. |
| $50-65/share | $5.6-7.3B | $30-38B | 4-6x | Good entry, not 10x but strong returns. |
| $80-100/share | $9-11B | $30-38B | 3-4x | Decent if you believe the mega bull case. |
| $118/share (today) | $13.3B | $30-38B | 2-3x | Already prices in significant growth. |
The 52-week low is $28.64. Tower traded at this level just 12 months ago. In a semiconductor downturn or broader market correction, these levels are reachable again.
At $35/share ($3.9B market cap), Tower would trade at:
At $35, with a bull case target of $30-38B, the math gives 8-10x. The probability of reaching $35 in a semiconductor downturn is moderate -- the stock was there 12 months ago. When it gets there, the risk/reward is exceptional because the floor ($35-55) overlaps with or is above the entry price.
| Capex intensity | Tower's capex is $400-450M/yr. TSMC's is $30B+. Leading-edge fabs cost $20B+ each. Tower's specialty fabs cost $300-500M to upgrade. The capex burden is orders of magnitude lower, which means Tower can self-fund growth. |
| Technology stickiness | Analog/mixed-signal processes take 2-3 years to qualify at a new foundry. Once a customer qualifies a process at Tower, switching costs are enormous. Leading-edge digital designs can be re-targeted more easily (TSMC to Samsung, etc.). Analog is "stickier." |
| Process longevity | A 130nm analog process designed in 2010 is still relevant and revenue-producing in 2026. There is no Moore's Law pressure to shrink. Tower's processes have 10-20 year revenue tails. Leading-edge processes are replaced every 2-3 years. |
| Diversification | Tower serves automotive, industrial, consumer, communications, medical, aerospace. No single end market dominates. Leading-edge foundries are increasingly dependent on AI/smartphone cycles. |
| Supply chain resilience | Tower's fabs are in Israel, USA, Japan, Italy. Geographically diverse and largely outside the Taiwan/China risk that dominates leading-edge production. |
| Bull Case (2030) -- Specialty AI Foundry Leader | $30-38B ($270-340/share) |
| Base Case (2030) -- Steady Specialty Growth | $18-25B ($160-225/share) |
| Bear Case -- Cycle Downturn / Multiple Compression | $5-8B ($45-72/share) |
| Floor (Asset-Based) | $35-55/share |
| 10x Entry Zone | $30-45/share ($3.4-5B market cap) |
| Current Price | $118.54 |
| 52-Week Low | $28.64 |
| 52-Week High | $149.57 |
| Intel's 2022 Bid | $53.00 |
| Analyst Mean Target | $159.29 |
Tower Semiconductor is a fundamentally excellent company. The balance sheet is a fortress, the technology portfolio is directly aligned with AI infrastructure demand (silicon photonics + power management), the management has been disciplined with capital allocation (paying down debt, investing in 300mm capacity, minimal dilution), and the competitive moat in specialty analog/mixed-signal processes is deep and sticky.
The problem is price. At $118.54, the stock has already repriced from "boring Israeli foundry" ($28 a year ago) to "AI infrastructure play" ($118 today). The 10x opportunity exists, but it existed at $28-$45, not at $118. At today's price, the upside to the bull case ($270-340) is 2-3x, which is good but not exceptional. The downside to the bear case ($45-72) is 40-60%, which is painful.
Our 10x entry zone is $30-45/share. At those levels:
When might we get there? Semiconductor downturns happen every 3-4 years. Tower dropped from $53 to $21 during the 2022-2023 downturn. A macro recession, AI demand pause, or sector rotation could bring the stock back to $40-60. That is when this analysis becomes actionable.
Key catalyst to monitor:
Compared to CORZ: Tower is the anti-CORZ. Clean balance sheet vs negative equity. Diversified customers vs 100% concentration. Consistent profitability vs cash burn. Moderate beta (0.94) vs extreme beta (6.9). Established specialty moat vs unproven pivot. The tradeoff: Tower is fairly valued today while CORZ offers more explosive upside if the pivot works. Leopold holds both -- Tower as the "safe" AI infrastructure play, CORZ as the asymmetric bet.
Data sources: SEC EDGAR XBRL (CIK 0000928876), 20-F annual report (FY2024, filed 2025-04-30), yfinance, Leopold Aschenbrenner 13F (Situational Awareness LP). Analysis date: 2026-03-12.