Laser shortage threatens AI data centers as Marvell and Nvidia lock up supply

I went through over 50 earnings calls so you don't have to, and one thing has become crystal clear:

Data centers are barreling straight into a brutal laser shortage.

Why $AAOI & $SIVE represent the biggest asymmetric opportunity in the AI trade: https://x.com/i/article/2120029522414153728

Inside the AI Laser Shortage: Why $AAOI and $SIVE Are the Most Asymmetric Bets in the AI Trade

On November 2, Marvell ($MRVL) walked out and paid $115 billion in stock for Lumentum ($LITE). Days later, Coherent ($COHR) disclosed a $20 billion, five-year supply deal for its PhotonLink platform and reminded analysts on its earnings call that it sells zero bare Indium Phosphide (InP) lasers to outside module makers. Back in March, Nvidia ($NVDA) had already wired $2 billion into Lumentum ($LITE) to lock up ultra-high-power laser capacity before anyone else could touch it.

Why are the biggest silicon designers on earth suddenly hoarding laser fabs?

Because silicon cannot emit light. Silicon has an indirect bandgap. TSMC ($TSM), Tower Semiconductor ($TSEM), and GlobalFoundries ($GFS) can etch world-class optical modulators and waveguides into silicon wafers, and Tower Semiconductor ($TSEM) and GlobalFoundries ($GFS) are both tripling their silicon photonics (SiPho) capacity into 2027 with order books completely spoken for. None of those silicon wafers work without an external Indium Phosphide continuous-wave (CW) laser firing O-band (1310nm) photons (the sweet spot where chromatic dispersion drops near zero across short-reach fiber and silicon waveguides) into the chip.

Right now, two physical walls, grid power starvation and laser facet melting, are colliding inside AI data centers, breaking global InP laser supply through 2028 with meaningful greenfield relief locked out until 2029. With Marvell ($MRVL) locking up Lumentum ($LITE) and Coherent ($COHR) hoarding its internal dies, non-Marvell ecosystems including Broadcom ($AVGO), AMD ($AMD), Astera Labs ($ALAB),GlobalFoundries ($GFS), and Jabil ($JBL) face an acute merchant laser shortage. That scarcity creates an asymmetric setup in two overlooked Western III-V fab owners: Applied Optoelectronics ($AAOI, ~$4.42B cap / ~$4.0B EV), scaling domestic U.S.laser epitaxy and 1.6T transceiver output toward a $5B+ run-rate, and Sivers Semiconductors ($SIVE.ST, ~$540M USD EV), an unencumbered Scottish merchant III-V foundry arming the non-captive ecosystem ahead of an H1 2027 U.S. Nasdaq dual listing.

Data Centers Are Out of Power, and Legacy Optics Torch Too Many Megawatts

To understand why optical architectures are changing so violently in 2027, follow the electricity inside a 100,000-GPU cluster.

Data centers are starved for utility hookups. GE Vernova ($GEV) sits on a $200 billion power backlog,while Vertiv ($VRT) saw organic cooling and power orders double year-over-year ahead of the 2027 shift to 800V DC racks. Inside a power-capped building,every watt burned moving data between racks is a watt stolen from revenue-generating GPUs.

A legacy DSP-retimed 1.6-terabit (1.6T) optical transceiver burns 25 to 30 watts, or roughly 15 to 18 picojoules per bit. Half of that electricity gets torched inside a tiny Digital Signal Processor (DSP) chip just cleaning up electrical signals across copper board traces. Multiply that across half a million optical links in a frontier cluster, and you are burning 12 to 20+ megawatts of scarce grid capacity just conditioning signals.

Hyperscalers are not abandoning pluggable form factors; instead, they are stripping out the power-hungry DSP for rack-to-rack networking while moving optical engines closer to the silicon for ultra-dense pods:

- Strip Out the Module DSP Now: Linear Pluggable and Linear Receive Optics (LPO and LRO) eliminate or half-retire the transceiver DSP inside pluggable modules, cutting module power in half right away.

- Move Optics Onto the Package Next: Near-Packaged Optics (NPO) and Co-Packaged Optics (CPO) place the optical engine next to or directly onto the switch and GPU substrate,dropping interconnect energy all the way down to 3 to 5 picojoules per bit.

Scale-Out vs. Scale-Up: Sizing the Two-Stage Rocket for InP Lasers

Most equity screens still model AI optics purely around scale-out networking, which connects racks to top-of-rack and leaf-spine Ethernet or InfiniBand switches across 50 meters to 2 kilometers. Scale-out alone is already exploding: AI optical transceiver sales jumped 60% to $26 billion in 2026, with 800G+ module shipments surging 2.6x past 63 million units.As clusters migrate from 4-lane 800G to 8-lane 1.6T pluggables across 2027 and 2028, scale-out requires roughly 2 to 3 transceivers per accelerator, translating to 16 to 24 discrete 70mW to 100mW EML or CW lasers per GPU, or an annual merchant and captive demand pool of 350 million to 500 million+ scale-out laser diodes (~$1.8B to $2.5B in bare chip value supporting a $30B+ module market).

Yet scale-out is only the opening act. Copper is hitting a hard physical wall inside scale-up networking, the memory-semantic fabric linking GPUs directly to other GPUs within a multi-rack domain, where bandwidth requirements sit 10x higher than scale-out. At the 2026 OCP Summit, Wiwynn (6669.TW) and privately held Ayar Labs demonstrated that once a 1,024-ASIC pod hits 3 megawatts of silicon power, copper cannot physically span the multi-rack scale-up domain. Their production tray uses 16 front-panel pluggable External Light Source (ELSFP) laser modules for every two ASICs.

Run the pod math on that scale-up layout:

- Per Chip: 8 high-power CW laser modules (typically housing 8- or 16-wavelength laser arrays, or 64 to 128 laser channels per accelerator)

- Per Rack: 512 laser modules

- Per 16-Rack Pod: 8,192 laser modules (before counting a single top-of-rack scale-out switch)

Corning ($GLW) expects scale-up pods to multiply fiber density tenfold from 16 to 160 fibers per accelerator by 2028, while Aehr Test Systems ($AEHR) confirmed 2027 and 2028 processors embed 4 to 8 optical I/O chiplets per package across a 25-million-unit processor market. Even if optical NPO and CPO penetrate just 15% to 25% of that 25-million-unit accelerator market across 2028 and 2029, at 32 to 64+ optical channels per package, scale-up creates an incremental 200 million to 400 million+ ultra-high-power CW laser channels. Because scale-up channels command $5 to $12+ per high-power CW die or array channel (and $45 to $100+ per multi-wavelength bar), scale-up adds another $1.5B to $3.5B+ in pure bare-laser silicon value on top of scale-out, effectively tripling wafer area consumption before decade-end.

How Pulling Lasers Off Hot GPUs Destroys Wafer Output

Shifting to NPO and CPO solves the cluster power problem, but it triggers a brutal physics problem for wafer supply.

Lasers hate heat, and hyperscalers hate scrapping a $40,000+ GPU tray over a dead $50 laser diode. You cannot bolt an InP laser directly onto a 1,000-watt GPU burning at 90 degrees Celsius without cooking it and creating an unserviceable single point of failure.So NPO and CPO designs pull the laser off the hot chip package and place it on the cool, hot-swappable front faceplate of the server as a pluggable External Light Source.

Pushing light from a front faceplate through connectors, fibers, and silicon waveguides creates heavy optical loss along the path. While a traditional scale-out pluggable transceiver runs fine on a 70mW to 100mW laser diode, NPO requires 100mW to 200mW lasers, and CPO requires 300mW to 400mW workhorses pumping out 23.4 dBm of optical power at 50 degrees Celsius.

As MACOM's ($MTSI) CEO put it on his quarterly call,CW laser survival is won or lost on the facet, the hottest microscopic point where light exits the crystal.Push 400mW through a standard cavity and thermal stress melts the mirror in what engineers call catastrophic optical mirror damage (COMD).

To stop the mirror from melting, physicists have to widen and stretch the laser waveguide cavity by 3x to 6x. Look at what that geometry change does to a 4-inch Indium Phosphide wafer:

- Standard 70mW Scale-Out Laser: Yields roughly 8,500 gross dies per 4-inch wafer.

- High-Power 300mW to 400mW Scale-Up CPO Laser: Collapses to roughly 1,500 gross dies per 4-inch wafer.

Taiwan's LandMark Opto (3081.TW), whose Q3 net profit jumped 528% year-over-year at 58.6% gross margins, confirmed this geometry eats up to three times more substrate per unit.

You do not have to wait for 2028 scale-up CPO deployments to feel this wafer crunch. Doubling scale-out lane counts from 800G to 1.6T pluggables in 2027, ramping early 100mW to 200mW LPO/NPO links, and running 2028 CPO qualification lots are already cannibalizing the exact same 4-inch epitaxy chambers today.

Locked Out Until 2029: Why New Supply and Rival Physics Cannot Bail Out Hyperscalers

Hardware squeezes only pay equity investors if fast followers or alternative technologies cannot rush in and ruin pricing. Across the compound semiconductor chain, equipment backlogs, substrate rationing, and seven-month reliability tests lock greenfield supply out through 2028, pushing real merchant relief into 2029:

- 6-Inch Wafers Will Not Rescue 2027: Bears often argue a shift from 4-inch to 6-inch InP wafers will erase the die penalty. AIXTRON's (AIXA.DE) CEO killed that theory on his Q3 call:6-inch MOCVD reactor shipments are dropping from 30% of mix down to 25% in 2027, while 3-inch tools are doubling from 15% to 30%,specifically because the world is short 6-inch InP substrates.

- 100% Wafer Rationing Through 2028: Global InP substrate leader AXT Inc. ($AXTI) grew Q3 sales 162% year-over-year to $73.4M, guided Q4 gross margins past 60% on a $500M+ binding backlog, and confirmed every customer is on strict allocation through all of 2028.Sumitomo Electric's (5802.T) infocommunications profit nearly tripled on the exact same shortage.

- Sold-Out Fab Tools and 5,000-Hour Burn-In Push Greenfield Supply to 2029: AIXTRON (AIXA.DE) MOCVD reactor lead times sit at 12 months (customer prepayments up 7x to EUR 309M), and Veeco ($VECO) has run out of factory floor space for its 2027 Lumina InP MOCVD and Spector facet-coating lines. Add cleanroom setup, epitaxy tuning, and mandatory 5,000-hour Telcordia reliability testing (~7 months of nonstop burn-in), and any greenfield fab ordered today cannot ship qualified commercial volume until early 2029.

- Delayed Western Challengers: MACOM ($MTSI) will not have internal 75mW to 100mW CW lasers in commercial volume until 2028, with greater than 100mW NPO/CPO devices pushed into 2029. WIN Semiconductors' (3105.TW) outsourced foundry customers face the same qualification clock, and Nokia's ($NOK) Arizona InP fab will not open until 2029.

- No Bailout From TFLN or Quantum Dots:Thin-Film Lithium Niobate (TFLN) lowers modulator loss slightly, but it is just a modulator material that still requires an external InP CW laser to function. Meanwhile, Quantum Dot (InAs/GaAs) lasers grown directly on silicon remain years away from clearing 5,000-hour Telcordia reliability at 85 degrees Celsius.

Checking the Hyperscale Receipt Book

A multi-year supply moat means nothing if downstream AI spending rolls over. Third-quarter filings and Anthropic's S-1 (proposed Nasdaq ticker:$ANT) prove downstream demand is accelerating,and that optical interconnects, not compute dies, are now gating deployments:

- Frontier Model & Cloud Commitments:Anthropic's public S-1 (ahead of its planned $ANT listing) revealed H1 2026 revenue jumped 14x YoY to $16.29B (with Q3 hitting ~$15.5B to $18.9B at ~85% inference gross margins and 900% enterprise net retention), backed by $131.7B in liquidity against $531.6B in compute purchase commitments and $321.4B in uncommenced data center leases. While frontier lab commitments flow directly into hyperscaler cloud backlogs rather than sitting purely on top of them, the underlying buildout is staggering:Microsoft ($MSFT) ($677B RPO, $50B+ quarterly CapEx), Amazon ($AMZN) ($578B backlog, $220B+ annual CapEx guiding 40%+ AWS growth through 2028), Alphabet ($GOOGL) (Google Cloud +103% to $30.7B,stepping up 2027 CapEx above ~$200B), and Meta ($META) ($825B in infrastructure commitments, with its Muse AI agent gated by interconnect supply) are all racing alongside newly public SpaceX ($SPCX), which is activating 220,000 Nvidia ($NVDA) GB300 GPUs per month toward 13+ gigawatts by late 2027.

- Binding Factory-Floor Bottlenecks: AMD ($AMD) guided 2027 revenue growth past 80% as MI450 Helios racks ship across ~3 gigawatts of deals, while Celestica ($CLS) guided 2027 sales to nearly double from $21B to $40B and explicitly called out optical transceivers and lasers as primary gating bottlenecks for rack deliveries. Upstream, optical packager Suzhou TFC (300394.SZ) (equipment prepayments +447%) confirmed CPO batch deliveries have started ahead of a 2027 ramp, and Tier-1 module maker Eoptolink (300502.SZ) saw non-DSP laser shortages force a 78x surge in supplier prepayments.

The Western Fab Scoreboard: Who Still Sells Unencumbered Lasers?

Look at the valuation and captivity scoreboard for who actually owns operational Indium Phosphide fabs, or explicitly refuses to build them:

- Lumentum ($LITE): Acquired by Marvell ($MRVL) for $115B (~16x to 18x EV/Sales), with $2B of ultra-high-power capacity already prepaid by Nvidia ($NVDA). Captive to Marvell ($MRVL) and Nvidia ($NVDA).

- Coherent ($COHR): Valued near $45B+ (~4.8x 2027E EV/Sales), locked into a $20B PhotonLink deal, and keeps 100% of its internal 4-inch and 6-inch bare InP laser dies captive for its own modules.

- LandMark Opto (3081.TW): Pure-play epi-wafer foundry printing 58.6% gross margins and trading near ~7.5x forward EV/Sales on surging substrate-constrained pricing.

- Astera Labs ($ALAB) & GlobalFoundries ($GFS): Astera Labs ($ALAB) (sampling NPO engines for H2 2027) and GlobalFoundries ($GFS) ($1.3B+ in SiPho customer prepayments) confirmed they will not manufacture InP lasers.

Non-Marvell ecosystems are forced to source independent lasers and transceivers externally.Applied Optoelectronics ($AAOI) and Sivers Semiconductors ($SIVE.ST) sit on the exact three attributes that create multi-bagger asymmetry during a structural semiconductor shortage, while offering distinct exposure across the scale-out and scale-upwaves:

1. Unencumbered Capacity Inside the 2027 to 2028 Scarcity Window: Anyone ordering MOCVD reactors today cannot ship qualified lasers until 2029. Applied Optoelectronics ($AAOI) and Sivers Semiconductors ($SIVE.ST) already own operating III-V cleanrooms and are bringing pre-qualified capacity expansions online across 2027 and 2028, right into peak hyperscale rationing.

1. Complementary Scale-Out Cash Flow & Scale-Up Torque: Applied Optoelectronics ($AAOI) monetizes the massive 2027 scale-outwave immediately via vertically integrated 800G and 1.6T pluggable transceivers (~85% to 90% of its near-term Data Center mix) while layering on greater than 60%-margin discrete 100mW to 400mW NPO/CPO lasers and greater than 40%-margin ELSFP modules for scale-up pods across five active customer engagements.Sivers Semiconductors ($SIVE.ST) is the pure-play merchant laser engine weighted heavily toward scale-up NPO/CPO and DSP-free LRO,where its 8- and 16-wavelength CW DFB laser arrays feed Ayar Labs' SuperNova ELSFP boxes, GlobalFoundries' ($GFS) Fotonix platform, and Jabil's ($JBL) pluggable modules.

1. Western Front-End Epitaxy at Disconnected Valuations: Against Lumentum's ($LITE) 16x to 18x buyout and Coherent ($COHR) and LandMark Opto's (3081.TW) 4.8x to 7.5x EV/Sales range, Applied Optoelectronics ($AAOI) sits at a ~$4.42B market cap / ~$4.0B EV (~$49/share) running domestic U.S. laser epitaxy in Texas, and Sivers Semiconductors ($SIVE.ST) sits at a ~5.37B SEK / ~$540M USD EV (~16.00 SEK/share) running a sovereign Scottish III-V fab in Glasgow. Both are bringing multi-billion-dollar revenue capacity online while still priced like legacy cyclical small caps.

Applied Optoelectronics ($AAOI): Domestic U.S. Vertical Integration at Scale

At a ~$4.42B market cap and ~$4.0B enterprise value (~$49/share, adjusting for $621.5M in cash against debt) after shedding more than 40% from its October highs, Applied Optoelectronics ($AAOI) got punished after Q3 earnings on two headlines: CenterPoint Energy ($CNP) delayed utility transformers for Phase 2 of its Sugar Land, Texas expansion by a few months, and Q4 Non-GAAP gross margin guidance (29.5% to 31.0%) came in below Street models.Looking under the hood of the November 5 filing shows a stark disconnect between temporary growing pains and contracted 2027 to 2028 earnings power.

Why the Bull Case Has Teeth:

- Surging Scale-Out Growth and $350M+ in Stranded 1.6T Orders: Q3 revenue jumped 40.6% QoQ (+127.5% YoY) to $269.9M, with Q4 guided up 49% QoQ at the midpoint to $385M to $420M ($0.48 to $0.63 Non-GAAP EPS).Anchor customer Microsoft ($MSFT) qualified Applied Optoelectronics' ($AAOI) 1.6T scale-out transceivers faster than third-party 1.6T DSP chip supply could keep up. Applied Optoelectronics ($AAOI) sits on $400M to $450M in firm 1.6T orders today, can only ship $50M to $70M in Q4 due to DSP lead times,and carries $350M+ of unshipped 1.6T backlog straight into Q1 2027 as a second DSP vendor qualifies.

- Multi-Billion-Dollar Hyperscale LTAs: Applied Optoelectronics ($AAOI) signed a three-year contract (LTA #1) through May 2029 worth $1.0B to $1.5B for at least 10% of its anchor customer's 1.6T volume. Better yet, CEO Dr.Thompson Lin disclosed on the Q3 call that LTA #2 with a second Tier-1 hyperscaler is in final contract redlines, expected within weeks, and sized at roughly 5x LTA #1 ($5.0B to $7.5B+) across 800G modules, discrete lasers, and next-gen assemblies.

- Proprietary Texas MBE Epitaxy & Locked-In Wafer Supply: While most rivals wait 12+ months on backlogged AIXTRON (AIXA.DE) and Veeco ($VECO) MOCVD reactors, Applied Optoelectronics ($AAOI) scales its domestic Sugar Land fab using custom-engineered Molecular Beam Epitaxy (MBE) paired with automated laser-to-fiber alignment. Historically,photonics rivals avoided MBE on InP due to slower chamber throughput and solid-source phosphorus handling complexity, but Applied Optoelectronics ($AAOI) spent two decades perfecting multi-wafer production MBE,delivering atomically sharp quantum-well interfaces for high-temperature CW lasers while sidestepping the global MOCVD tool bottleneck entirely. To feed those chambers, Chinese wafer producer Yunnan Germanium (002428.SZ) filed regulatory notice of a binding five-year deal supplying Applied Optoelectronics ($AAOI) with 900,000 4-inch InP wafers (15,000/month,including a $194.4M binding 2027 commitment),supplemented by five Japanese and German substrate vendors capable of backstopping critical Western runs. At scale, Applied Optoelectronics ($AAOI) can print 720M to 1.44B laser diodes a year.

- 1M/Month Scale-Out Target + >60% Margin Scale-Up Lasers: While December 2026 module capacity ticked down from 650k to greater than 500k units/month as Texas Phase 1 opens, management reaffirmed its year-end 2027 target above 1,000,000 scale-out transceivers/month ($471M/month, or a $5.65B annualized Data Center transceiver target).Layered directly on top of that transceiver base is Applied Optoelectronics' ($AAOI) unmodeled scale-up call option: the company is already shipping evaluation quantities of 300mW to 400mW CPO lasers across five active NPO/CPO customer engagements, carrying greater than 60% gross margins on bare CPO laser chips and greater than 40% gross margins on ELSFP and CPO modules. Even allocating 15% to 20% of its Texas InP wafer output to merchant scale-up CPO/NPO lasers and ELSFP assemblies represents $500M to $1.0B+ in high-margin incremental revenue by 2028.

The Real Warts You Have to Underwrite:

- 41% of Q3 Sales Came From Cable TV Gear:Data Center generated $155.6M in Q3 (57.7% of sales), while legacy CATV gear generated $110.9M (41.1%). You cannot slap an AI foundry multiple on a ~$440M annualized cable business (~1.0x EV/Sales value).

- Module Assembly vs. Foundry Margin Mix & Negative GAAP EBIT: Headlines cheered $6.3M in Q3 GAAP net income, missing that GAAP operating income was -$1.44M ($76.16M gross profit minus $77.60M OpEx, up 51% YoY),lifted into the black by $10.1M of interest and other income. Furthermore, scale-out transceiver assembly carries ~30% to 38% gross margins,roughly half the ~58% to 60% gross margins of pure-play laser foundries like LandMark Opto (3081.TW), meaning Applied Optoelectronics' ($AAOI) module revenue deserves a lower EV/Sales multiple than its greater than 60%-margin discrete scale-up lasers.

- LTAs Are Capacity Ceilings With Baked-In Price Cuts: CFO Stefan Murry noted LTA #1 bakes in annual price reductions over three years and includes zero upfront customer CapEx funding (though customer co-investment remains on the table specifically for newer CPO lines).As Applied Optoelectronics ($AAOI) investors learned in 2017, hyperscale optical LTAs often act as capacity reservation ceilings rather than strict take-or-pay minimums. Q3 Non-GAAP gross margin (28.7%) also absorbed lower-margin 400G backfill ($93.4M) and a ~250 bps non-cash contra-revenue drag from vesting hyperscaler warrants.

- Working Capital Burn & Residual China Exposure: Funding plant commitments, $350M+ in 1.6T WIP, and $194.4M in wafers requires heavy working capital before late-2027 free cash flow inflects, making Applied Optoelectronics' ($AAOI) $621.5M cash pile and new $200M+ credit line vital bridges. And while front-end laser fabrication is 100% U.S.-based in Sugar Land,~70% of back-end module assembly still runs in Taiwan and Ningbo, China, and heavy reliance on Yunnan Germanium (002428.SZ) leaves Applied Optoelectronics ($AAOI) exposed to margin friction if Beijing restricts Indium or Germanium exports before its five Japanese and German substrate suppliers and Texas Phase 2 automated lines shoulder full volume.

Stress-Tested 2027 to 2028 Valuation Scenarios (vs. ~$49/Share | ~$4.42B Cap | ~$4.0B EV Today):

Because scale-out transceiver modules carry ~35% to 38% target gross margins while merchant scale-up CPO/NPO lasers carry greater than 60% gross margins, rigorous underwriting requires a sum-of-the-parts (SOTP) EV/Sales and forward P/E framework across ~90M to 96M fully diluted shares (valuing CATV at 1.0x EV/Sales, separating scale-out module multiples from discrete scale-up laser multiples, and adjusting for net cash and working capital draws):

- Thesis-Breaker Downside (LTA #2 Collapses + Working Capital Crunch): If LTA #2 fails to sign, DSP and transformer delays stall Data Center revenue at ~$1.4B, and warrant/price-cut drag pins gross margins below 30% while working capital eats the cash buffer, the market will value Applied Optoelectronics ($AAOI) like a cyclical EMS assembler at 1.2x Data Center EV/Sales (~8x depressed EPS). That breaks the momentum story and pulls the stock down to ~$24 to $30/share (-40% to -50% downside).

- Scenario 1, Conservative Execution Haircut (50% Haircut to 1M/Month Target):Transformer and DSP delays cap Applied Optoelectronics ($AAOI) at $2.2B to $2.5B in scale-out Data Center sales (+ $450M CATV) with minimal CPO contribution and consolidated gross margins stalling near 33% to 34%. Valuing transceiver modules at a conservative 2.2x to 2.6x EV/Sales (~11x to 13x EPS, well below Coherent's ($COHR) 4.8x to reflect lower module margins) plus CATV at 1.0x yields a $5.7B to $7.3B equity value ($63 to $81/share),or +28% to +65% upside even if management misses its expansion target by half.

- Scenario 2, Base Case (70% Scale-Out Conversion + Initial Scale-Up Laser Mix):Applied Optoelectronics ($AAOI) converts roughly 70% of its late-2027 transceiver target plus initial NPO/CPO discrete laser ramps ($3.5B to $4.0B Data Center total + $450M CATV), lifting Non-GAAP gross margins past 37% to 40% and driving $8.50 to $10.50+ in run-rate EPS. Blending 3.0x to 3.5x EV/Sales on scale-out modules with 5.5x to 6.5x on greater than 60%-margin scale-up lasers (~15x to 17x blended EPS, still a discount to Coherent's ($COHR) 4.8x sales) values Applied Optoelectronics ($AAOI) at $12.2B to $16.8B ($135 to $187/share), or +175% to +280% upside (2.8x to 3.8x).

- Scenario 3, Bull Case (Full 1M/Month Scale-Out + Multi-Customer Scale-Up CPO Ramp):Applied Optoelectronics ($AAOI) hits its full $5.0B to $5.65B scale-out run-rate plus $600M+ in high-margin discrete CPO/ELSFP scale-up sales across multiple hyperscalers, pushing blended gross margins above 40% and run-rate EPS past $15.00. At a blended 4.2x to 5.0x Data Center EV/Sales (~18x to 20x EPS,matching Coherent ($COHR) and still less than one-third of Lumentum's ($LITE) buyout multiple), Applied Optoelectronics ($AAOI) commands a $22.0B to $28.7B valuation ($245 to $318/share), or +400% to +550% upside (5.0x to 6.5x).

Sivers Semiconductors ($SIVE.ST): The Last Independent Western Laser Foundry

While Applied Optoelectronics ($AAOI) captures near-term scale-out module dollars on its way into discrete scale-up lasers, Sivers Semiconductors ($SIVE.ST,thinly OTC: $SIVEF) is built purely to arm the merchant ecosystem with bare III-V laser arrays.Sitting on one of the last independent Indium Phosphide fabs in the West, Sivers Semiconductors ($SIVE.ST) is transitioning from NRE engagements into commercial product volume right ahead of a planned U.S. Nasdaq dual listing.

Where Valuation Stands Right Now:

Sivers Semiconductors ($SIVE.ST) is up nearly 4x over the past year, carrying a ~5.37B SEK (~$540M USD) market cap and roughly equivalent enterprise value at ~16.00 SEK/share (~$1.60 USD) on 277.9M SEK (~$28M USD) in trailing twelve-month revenue (Q2 sales were 53.8M SEK; adjusted EBITDA was -35.5M SEK). At ~19x trailing EV/Sales ahead of its November 26 Q3 report, European tech funds already recognize the scarcity story, meaning execution missteps are not free.

Why Ecosystems Need Sivers Across Both Scale-Out and Scale-Up:

- An Unencumbered Western InP Fab: CEO Vickram Vathulya controls a wholly owned,operating III-V InP fab in Glasgow, Scotland (formerly CST Global), backed by a five-year customer opportunity pipeline up 268% since year-end 2025 to $1.2B.

- A Two-Stage Scale-Out to Scale-Up Revenue Bridge: Sivers Semiconductors ($SIVE.ST) does not have to wait for 2028/2029 CPO pods to monetize Glasgow. Its near-term 2027 bridge targets DSP-free scale-out: 8-channel, 100mW O-band CW DFB laser arrays and Semiconductor Optical Amplifiers entering 800G/1.6T LRO pluggable beta builds with Jabil ($JBL) late in Q4 alongside a $3.4M SemiNex AI optical program (plus an $8.2M 2027 mmWave SATCOM production order from privately held ALL.SPACE). Its massive 2028 to 2029 torque sits squarely in optical scale-up,where Sivers Semiconductors' ($SIVE.ST) 8- and 16-wavelength CW-WDM laser arrays power Ayar Labs' SuperNova ELSFP external light sources and GlobalFoundries' ($GFS) Fotonix SiPho platform. Because a single scale-up GPU package consumes 4 to 8 multi-wavelength arrays (32 to 128 laser channels per chip vs. 8 channels in a scale-out pluggable),winning just a sliver of non-Nvidia merchant scale-up sockets can absorb Sivers Semiconductors' ($SIVE.ST) entire 100M-channel internal Glasgow expansion and 150M to 250M outsourced foundry channels.

- Navigating the Monolithic Array Yield Trap:Multi-wavelength 8- and 16-channel CW arrays carry a greater than 50% ASP and margin premium over $3.50 to $4.50 single scale-out diodes ($4.50 to $7.50 per high-end channel, or $45 to $100 per array). Historically, monolithic multi-wavelength DFB arrays suffer painful early yields because minor grating pitch drift across CW-WDM MSA wavelength spacing causes a single out-of-spec channel to scrap the entire bar. Sivers Semiconductors ($SIVE.ST) attacks this two ways: precision electron-beam/holo-grating uniformity in epitaxy to lift raw bar yield,paired with wafer-level etched-facet testing to screen out bad bars before incurring expensive back-end packaging costs. Proving it can hold 55% to 60% gross margins while scaling Glasgow 4x to 100M channels/year by Q4 2027 remains the core engineering test.

Stress-Tested 2028 Valuation Scenarios (vs. ~16.00 SEK / ~$540M USD EV Today):

Unlike Applied Optoelectronics' ($AAOI) mixed module model, Sivers Semiconductors ($SIVE.ST) is a pure-play merchant III-V semiconductor supplier targeting 55% to 60% gross margins, making Coherent ($COHR) (4.8x EV/Sales), LandMark Opto (3081.TW) (7.5x EV/Sales), and Lumentum ($LITE) (16x to 18x buyout) true margin-adjusted peers.However, because greenfield MOCVD supply begins arriving in 2029, disciplined 2028 exit multiples must bake in a mid-cycle compression discount unless locked under multi-year LTAs. Modeling across ~355M pro-forma diluted shares (including a 5% to 8% bridge/U.S. listing buffer):

- Thesis-Breaker Downside (Array Yield Stalls / Telcordia Slippage): If Sivers Semiconductors ($SIVE.ST) stumbles on 5,000-hour Telcordia reliability above 100mW or monolithic 8-channel grating yields stall below commercial thresholds,forcing a dilutive equity raise while delaying volume POs into 2028, the 19x trailing scarcity premium evaporates. Re-rating back toward R&D/SATCOM sum-of-the-parts (~$180M to $230M EV) cuts the stock to ~5.50 to 7.00 SEK/share (-55% to -65% downside).

- Scenario 1, Conservative Execution Haircut (Outsourced Foundries Slip to 2029; Glasgow Only): Outsourced foundry partners hit industry-wide qualification delays through 2028 (a 75% total capacity haircut), leaving Sivers Semiconductors ($SIVE.ST) selling only 80M to 100M internal Glasgow channels across LRO scale-out and early NPO scale-up by late 2028 at $4.50 ASP. That still generates $360M to $450M USD (~3.6B to 4.5B SEK) in ~55%+ gross margin photonics run-rate sales. Applying a mid-cycle-discounted 3.5x to 4.2x EV/Sales multiple (below Coherent ($COHR)) values Sivers Semiconductors ($SIVE.ST) at $1.26B to $1.89B USD (35 to 53 SEK/share, or ~$3.55 to $5.30 USD), delivering +120% to +230% upside (2.2x to 3.3x) off Glasgow alone.

- Scenario 2, Base Case (50% Total Capacity Utilization + Scale-Up NPO Ramp): Converting half of total 2028 capacity (125M to 175M channels across Glasgow and initial outsourced foundry lines at $4.50 to $5.25 ASP as 8/16-channel scale-up arrays mix in) yields $560M to $920M USD (~5.6B to 9.2B SEK), pricing today's EV at just 0.6x to 1.0x run-rate sales.Applying 4.5x to 5.8x EV/Sales (discounting LandMark Opto's (3081.TW) 7.5x to account for 2029 supply arrivals) drives a $2.5B to $5.3B USD valuation (71 to 150 SEK/share, or ~$7.10 to $15.00 USD), or +345% to +835% upside (4.4x to 9.4x).

- Scenario 3, Bull Case (Full 250M to 300M Scale-Up & Scale-Out Ramp / Strategic M&A Takeout): Full utilization backed by multi-year hyperscale NPO/ELSFP scale-up LTAs generates $1.25B to $1.80B USD in high-margin run-rate revenue (factoring in modest volume ASP step-downs as 2029 approaches). At a strategic M&A takeout or scarcity multiple of 5.5x to 7.0x EV/Sales, still well below half the multiple Marvell ($MRVL) paid for Lumentum ($LITE), Sivers Semiconductors ($SIVE.ST) commands a $6.9B to $12.6B USD valuation (194 to 355 SEK/share, or ~$19.40 to $35.50 USD), representing 12x to 22x upside as the last independent Western photonics target on the board.

Two Major H1 2027 Catalysts (Customer Prepayments + U.S. Nasdaq Dual Listing):

- Eliminating Dilution With Customer Prepayments: Running at roughly -$3.5M USD in quarterly adjusted EBITDA burn before CapEx, Sivers Semiconductors ($SIVE.ST) cannot self-fund $30M in Glasgow tool expansion plus InP wafer WIP off its existing cash buffer and $28M in trailing revenue.Against a ~$540M USD cap, a standard facility or equity raise is a manageable 5% to 8% dilution hit, but if Sivers Semiconductors ($SIVE.ST) pairs H1 2027 commercial NPO/LRO awards with upfront customer capacity prepayments (matching GlobalFoundries ($GFS), Coherent ($COHR),and Veeco ($VECO)), dilution risk drops toward zero.

- Crossing Onto U.S. Nasdaq: Most U.S. semi funds cannot buy Stockholm small caps or illiquid OTC shares. Sivers Semiconductors ($SIVE.ST) shifted its reporting calendar in July to meet U.S. PCAOB standards, and at its October 22 EGM, shareholders appointed Ernst & Young for SEC-grade audits and approved a U.S.-aligned option plan ahead of a targeted H1 2027 U.S. Nasdaq dual listing, unlocking U.S.institutional capital to price Sivers Semiconductors ($SIVE.ST) directly alongside AXT Inc. ($AXTI) and Coherent ($COHR).

Three External Macro Risks That Can Humble Both Stocks

Even if both management teams hit every internal milestone, three industry and geopolitical traps require close attention:

- Japanese III-V Pivots and Merger Risk:Legacy laser makers are shutting down old 25G/100G C-band/O-band telecom lines. If Sumitomo Electric (5802.T), Mitsubishi Electric (6503.T), and Furukawa Electric (5801.T) retool telecom MOCVD chambers into high-power 1310nm O-band CW data center lasers faster than expected in 2027 and 2028, or if antitrust regulators block Marvell ($MRVL) from closing on Lumentum ($LITE), merchant laser pricing power will cool off early.

- Unfinished Grid Permitting Reform (S. 5653):Federal permitting relief is not settled law yet.On November 9, the U.S. Senate voted 73 to 25 on a procedural cloture motion simply to open floor debate on S. 5653 (the BAAJA permitting bill). It still faces Senate amendment fights and stiff lame-duck House pushback over NEPA judicial limits and Tribal consultation rules before January. Even if enacted, S. 5653 directs FERC within 90 days to force data centers to fund 100% of incremental grid upgrades (~$35B over a decade), post upfront financial guarantees,and let states cap data center power draw.Power stays scarce and expensive either way,reinforcing low-watt optics while risking site delays.

- The Iran War, Strait of Hormuz, and Rate-Driven De-Grossing: There is a reason the Nasdaq just slid into correction territory and high-beta optical stocks got hit alongside it. The 2026 Iran conflict and shipping disruptions through the Strait of Hormuz triggered one of the sharpest energy supply shocks in decades,keeping U.S. gasoline above $4 a gallon for 113 straight days and pinning the 10-year U.S.Treasury yield near 4.85% to 4.88% on sticky 3.7% inflation and December Fed rate-hike fears. Even though Brent crude just pulled back to ~$93 a barrel on Qatar-brokered peace proposals and cooler U.S. strike rhetoric, attacks on vessels near the Strait of Hormuz have not stopped. Any breakdown in talks that sends Brent back above $100 and pushes the 10-year yield past 5% will trigger immediate quant pod liquidation across mid- and small-cap semiconductors regardless of how strong their 2027 contract backlogs look.

The 24-Month Execution Roadmap: Quarter-by-Quarter Pass/Fail Gates

Instead of trading macro noise, track each company independently against hard pass/fail gates over the next eight quarters:

1. Late November through Year-End 2026 (Immediate Ink & Beta Builds)

- Applied Optoelectronics ($AAOI) Gate: Watch for an 8-K announcing formal signature of LTA #2 ($5.0B to $7.5B+). If it slips past Q1 without explanation, cut exposure; if signed with firmer pricing protections than LTA #1, the stock re-rates fast. Also verify December activation of Sugar Land Phase 1 (210,000 sq ft), pushing scale-out module capacity past 500,000 units/month.

- Sivers Semiconductors ($SIVE.ST) Gate:Reports Q3 results on November 26, 2026. Look for confirmation that Jabil's ($JBL) DSP-free LRO scale-out beta builds are kicking off on time, photonics pipeline growth above $1.2B,cash runway disclosures, and concrete SEC filing steps with EY toward the H1 2027 U.S.Nasdaq listing.

2. Q1 2027 (Clearing Stranded Backlog & Full-Year Audits)

- Applied Optoelectronics ($AAOI) Gate: On the late-February Q4 print, two metrics dictate everything: qualification of the second 1.6T DSP chip supplier (unlocking the $350M+ stranded 1.6T backlog) and Q1 Non-GAAP gross margin guidance clearing 32.5% to 34.0%. Staying stuck below 31% would mean warrant drag and LTA #1 price cuts are eating the 1.6T mix shift.

- Sivers Semiconductors ($SIVE.ST) Gate:Reports full-year/Q4 results on February 25,2027, delivering its first audit checkpoint under EY alongside initial beta and multi-channel grating yield readouts from Jabil ($JBL) and SiPho partners.

3. Q2 2027 (The Prepayment Test & U.S. Nasdaq Debut)

- Applied Optoelectronics ($AAOI) Gate: Track customer qualifications on its greater than 60%-margin discrete 100mW to 400mW scale-up NPO and CPO lasers across its five active target engagements, alongside initial LTA #2 order flow.

- Sivers Semiconductors ($SIVE.ST) Gate:Make-or-break commercial window. Engineering programs with Ayar Labs, GlobalFoundries ($GFS), SemiNex, and Jabil ($JBL) must convert into binding multi-year scale-up NPO and scale-out LRO volume purchase orders (ideally paired with upfront customer cash deposits to fund Glasgow cleanly) alongside going live on U.S. Nasdaq.

4. Q3 2027 (Power Energization & GAAP Operating Profit)

- Applied Optoelectronics ($AAOI) Gate:CenterPoint Energy ($CNP) must energize the delayed utility transformers at Sugar Land Phase 2. Simultaneously, Non-GAAP gross margins need to reach management's 37.0% to 40.0% target range, flipping GAAP operating income solidly positive without relying on interest income.

- Sivers Semiconductors ($SIVE.ST) Gate: Tool installation and cleanroom qualification checkpoints at Glasgow ahead of Q4 commissioning, plus initial 5,000-hour reliability readouts from outsourced foundry partners for 2028 capacity.

5. Q4 2027 through 2028 (Full Commercial Scale)

- Applied Optoelectronics ($AAOI) Gate:Scaling past 1,000,000 monthly scale-out transceivers, ramping merchant CPO/ELSFP scale-up modules, absorbing the $194.4M Yunnan Germanium (002428.SZ) wafer ramp alongside Japanese and German second sources smoothly, and inflecting to strong positive free cash flow.

- Sivers Semiconductors ($SIVE.ST) Gate:Commissioning the $30M, 100M-channel Glasgow expansion right as Astera Labs ($ALAB), GlobalFoundries ($GFS), Ayar Labs,and Fabrinet ($FN) ramp scale-up NPO and ELSFP volume into hyperscale pods (alongside Sivers Semiconductors' ($SIVE.ST) $8.2M ALL.SPACE SATCOM production ramp),followed by 150M to 200M outsourced foundry channels coming online across 2028 and 2029.

Three Macro Dials to Check Every Quarter:

- Substrate & Tool Checks ($AXTI & AIXA.DE):Monitor quarterly calls from AXT Inc. ($AXTI) and AIXTRON (AIXA.DE) to confirm 6-inch InP wafers stay tight and Japanese telecom MOCVD lines are not flooding 1310nm O-band CW supply ahead of schedule.

- Antitrust Dockets ($MRVL / $LITE): Watch MOFCOM and global regulatory reviews on the Marvell ($MRVL) and Lumentum ($LITE) merger ahead of its targeted H1 2028 close.

- Strait of Hormuz & 10-Year Yields: Watch whether Qatar-mediated U.S.-Iran diplomacy durably reopens Strait of Hormuz tanker traffic and pulls the 10-year Treasury yield down off ~4.85%; a sustained oil unwind removes the single biggest macro headwind compressing high-growth optical multiples.

When industry insiders spend $115 billion buying laser fabs and prepaying years of wafer output in cash,they are telling you exactly where the next multi-year choke point sits. Silicon photonics cannot scale without Indium Phosphide, 400mW cavity physics actively shrink wafer yields as speeds climb, and seven-month Telcordia burn-in clocks prevent fast followers from spoiling pricing before 2029. Against that backdrop, Applied Optoelectronics ($AAOI) and Sivers Semiconductors ($SIVE.ST) stand out not because they are flawless businesses, but because the gap between their contracted 2027 to 2028 capacity and their current enterprise values is wide enough to absorb real execution friction and still deliver multi-bagger returns. In a cornered component market, owning the last unencumbered fabs in the West is simply good math.