I started watching the $SIVE and $JBL connection back in March, before any of this was public, just breadcrumbs. It’s developed a lot faster than I expected since then. What started as one relationship has turned into six other manufacturers following the same path, a real production timeline, and what I think is the clearest near term catalyst in the entire thesis. This is meant to be everything you need to know in one place: Jabil start to finish, where the other six actually stand, why the standard product approach matters more than people give it credit for, what the broader pluggables market and shortage actually look like, and what I think happens to the stock as more of these convert.

The Jabil Relationship, Start to Finish

On April 15, 2026, Sivers announced a collaboration with Jabil. Jabil is building a 1.6T linear receive optical (LRO) transceiver module, a pluggable that plugs into a GPU or switch board and lets AI data center systems talk to each other over fiber at very high speed, using Sivers’ DFB lasers as the light source inside it. Sivers doesn’t build the finished module, Jabil does. Sivers supplies the laser chip that goes inside it. The pitch was energy efficiency, Sivers said the design delivers the same bandwidth at roughly 2.5x lower power than existing alternatives, which matters because power is now as much a constraint on AI data centers as compute is.

Here’s where things actually stand as of the Q2 call:

• Alpha builds, already done: early sample lasers went into Jabil’s design to prove the basic concept works

• Beta builds, happening now through Q4 2026: a more refined version meant to catch whatever issues showed up in alpha before the design locks in

• Customer qual cycle comes next: the actual hyperscaler buying Jabil’s finished module runs its own formal tests before approving it for real volume

• Initial production orders expected H1 2027, with the real ramp planned H2 2027

Using a simple four-stage framework, Jabil has completed the alpha stage and is moving into beta, which means two more stages have to happen before real revenue shows up. That’s honestly what got me excited about this back in March when it was still just breadcrumbs, and it’s developed exactly on the timeline management laid out since

What Actually Happened to the Stock

The Jabil news hit on April 15. The very next day, April 16, Sivers also announced a 125 million SEK directed share placement to a group of institutional investors, and separately announced it was evaluating a potential Nasdaq dual listing. Three real catalysts landed inside of about 24 hours.

That makes it hard to isolate “the Jabil reaction” as one clean number, the market was digesting all three at once.

Who’s Actually Buying This

Neither Sivers nor Jabil has named the hyperscaler on the other end of this. That’s normal at this level of the supply chain, Sivers sells a component to Jabil, Jabil sells a finished module to whichever hyperscaler is deploying it, and end customer names usually don’t get disclosed this early, if ever, for competitive reasons. The realistic universe here is the handful of hyperscalers actually building AI clusters at a scale that needs 1.6T interconnects, which today is still a short list.

The Other Six, Where Each One Actually Stands

Now there are six other manufacturing companies following the same path Jabil did. “Six companies” doesn’t mean they’re all in the same spot though, so here’s where each one actually stands.

Three of them are in alpha sample evaluation. Real sample lasers are in their hands right now, being tested in their own systems, roughly where Jabil itself was before it moved into beta builds.

The other three are a step earlier, in technical engagement and supply assessment. No physical samples yet, still working out specs and whether Sivers can actually deliver the volume and timeline they’d need. That puts them behind the alpha group in the process.

Management also said some of these engaged manufacturers are capable of very rapid qualification and ramp, without saying which ones specifically. There are also more prospects beyond just these six that haven’t been named at all.

Realistic read: the three in alpha are the closer group, tracking maybe a stage behind Jabil. The three in technical engagement have more ground to cover before they’re even at that point. Not all six convert on the same timeline, and this is still a non-binding pipeline, so not all of them convert at all.

One more detail worth pulling from the actual roadmap slide: every other H1 2027 milestone box names the specific customer, Tier-1 Telco, Tachyon. The pluggables box just says “Pluggables,” not Jabil by name. That’s a small thing, but it’s consistent with more than one company showing up in that window rather than just Jabil, especially with three of the other six already in alpha sample evaluation and some of them capable of what management called very rapid qualification and ramp. I can’t confirm that’s what it means, Sivers hasn’t said so directly, it’s a reasonable read of how that box is labeled compared to every other one on the slide, not a confirmed fact.

Why the Other Six Are Probably Hyperscaler-Tied Too

For a manufacturer to seriously engage a new laser supplier around a 1.6T pluggable program, the supplier generally needs to clear two major hurdles: the product has to meet the technical requirements, and the supplier needs a credible path to volume. Sivers has said they have both, production-ready standard parts already qualified in Jabil’s design, and a new foundry partner bringing on capacity the CEO has directly said isn’t a bottleneck right now.

Given that, think about who’s actually asking for 1.6T pluggable transceivers in the first place. This isn’t a broad market, it’s specifically hyperscale AI data centers, that’s the exact framing Sivers used when they announced the Jabil collaboration. If a manufacturer is engaging Sivers about a 1.6T pluggable right now, the realistic buyer on the other end of that manufacturer’s module is very likely a hyperscaler, because that’s close to the only customer base actually deploying at that speed today.

This same pattern is playing out elsewhere in the ecosystem right now, not just with Sivers. Aeva is a good example, they recently launched their own Optical Connectivity business with a hyperscaler relationship already behind it. Real hyperscaler demand is what’s pulling new suppliers into this chain across the board right now.

Put those two things together and the technical progress suggests at least some of these relationships are getting close enough that another named customer announcement is plausible, but the timing ultimately depends on the customer’s willingness to disclose the relationship, not just how far along the technical work is. If one does get named, I’d expect it to look a lot like the Jabil announcement itself, and given who’s actually buying this category of product right now, there’s a real chance it’s tied to another high profile name too.

Timing a PR Like Jabil’s, Working Backward From the Actual Dates

Let’s actually build this out instead of just saying “soon.” Jabil’s real timeline: collaboration announced April 15, 2026, alpha builds complete by the Q2 call in August, call it about four months. Beta builds land in Q4 2026, then a qual cycle, then initial production orders H1 2027, ramp H2 2027.

The three companies currently in alpha sample evaluation are, per management’s own framing, roughly where Jabil was right before it moved into beta. Their engagement appears to have followed the April Jabil validation, although Sivers has not disclosed the exact start date for each relationship. If that’s roughly right, getting from a fresh start to that same alpha-complete position by the August call would put the window at four months or less, which would be the “production ready chip, customers can ramp quickly” effect potentially showing up.

If one of these three keeps pace with Jabil’s own cadence from this exact point forward, you’d expect something like beta-equivalent work late this year into early 2027, a qual cycle through Q1 or Q2 2027, and a shot at initial orders somewhere in the H1 to H2 2027 window, similar to Jabil or maybe a quarter or two behind it given the later start.

Why Standard, Not Custom, Is the Whole Ballgame

Sivers already has its 70mW and 100mW laser singles and arrays production ready, the exact tier used in Jabil’s design, and it’s a standard part, not something custom built just for them. That’s part of why the ripple effect to the other six looks real instead of hopeful. Sivers has said outright they’re shifting from custom products to standard ones specifically because standard products move faster to production and reach more customers. A new manufacturer doesn’t need Sivers to design them a laser from scratch, they need Sivers to hand them one that’s already qualified and already in production. That’s a fundamentally shorter path than what a brand new component usually takes.

There’s a second wave forming behind this one too. In just the last seven months, an interim category has picked up real momentum, extra dense pluggables and pluggable CPO, sitting between traditional pluggables and full co-packaged optics. Sivers is now sampling the next tier up for that, 200mW laser singles and 100mW arrays, to get that ready the same way it got the 70/100mW tier ready for Jabil. If the fast movers in that group are actually fast, some of this could land alongside the pluggables wave in 2027 instead of years behind it.

The Three Waves

Here’s how I’d stack the whole roadmap. Wave one is traditional pluggables, Jabil and whichever of the other six convert first, this is the near term revenue while everything else matures. Wave two is XPO and pluggable CPO/NPO, the intermediate category that’s picked up momentum over the last seven months, using the 200mW tier Sivers is sampling now. Wave three is full co-packaged optics, GFS SCALE and Ayar Labs, still the biggest long term opportunity but the furthest out, more of a 2028 story. Pluggables are the bridge that gets Sivers real revenue while waves two and three mature behind it.

AAOI Is the Proof This Market Is Real, Not Hypothetical

If you want to know whether this demand is real or just a story, look at what’s already happening one rung up the supply chain. Applied Optoelectronics ($AAOI) reported Q2 2026 results in August, and on that call, CEO Thompson Lin said customer demand is running 20% to 40% higher than what they can currently supply, and that they’re hearing from big customers “almost every week, every month” asking them to speed up delivery. Management said flatly they have to be careful about taking too many new orders until capacity catches up.

CFO Stefan Murry put it even more bluntly on a separate call a couple weeks later, AAOI is effectively sold out through at least the second half of 2027 for certain products and has to avoid overcommitting capacity to new customers. One Yahoo Finance piece covering it framed it well: every order AAOI turns away is one a competitor takes instead.

Here’s the part that matters most for SIVE specifically. On that same AAOI call, Thompson Lin was asked about bottlenecks and said the laser is literally the biggest bottleneck in the whole transceiver business right now, not DSPs, not TIAs, the laser. That’s not me making the connection, that’s AAOI’s own CEO naming the exact node Sivers sits in. AAOI is expanding its own capacity aggressively just to keep up, including a major US manufacturing buildout, and even with that, they’re telling the market demand outstrips supply into 2027. That’s the environment a newly qualified laser supplier is stepping into.

Is Sivers Late to This?

The honest answer to “aren’t they late to pluggables” is that six months ago you could’ve made that case. Now there are six other manufacturers actively engaging, Jabil is already through alpha and into beta, and the laser itself is the acknowledged bottleneck across the industry per AAOI’s own commentary above. Late would mean entering a mature, fully qualified supply chain with little room for new suppliers. That’s not what we’re seeing. A 20-40% supply shortfall and a sold-out-through-2027 competitor one rung up the chain point the other way.

Worth noting too: this wasn’t a last-minute pivot chasing a hot narrative. In that same Q2 2026 shareholder letter, the CEO ties the current pluggables traction back to “our strategic move in Q3 2025, to support pluggables in addition to CPO.” That’s a deliberate decision made months before Jabil was ever announced, not a reaction to it.

The CEO has also said in a podcast that he expects pluggables to be a relevant product category for the next 10-plus years, not a short window that closes.

How Fast This Scales, and What the Revenue Shape Looks Like

Sivers isn’t designing a new laser for each new customer, it’s handing them a part that’s already production-ready and already validated in Jabil’s own design. That’s the actual reason management can say some of these engaged manufacturers are capable of what they called very rapid qualification and ramp. The bottleneck on a new engagement now is the customer’s own qualification cycle and how many relationships Sivers can service at once, not laser development time.

On whether more show up in Q3 or Q4: I’d expect the count to keep growing, though I’m not going to guess a specific number. The pattern so far is zero to six engaged manufacturers in about four months once Jabil validated the concept, and the CEO explicitly said there are more prospects beyond the six that haven’t even been named as engagements yet, “many other prospects continue to be identified” was the actual phrase. The realistic read for Q3 and Q4 is some of the three currently in technical engagement moving up into alpha, and possibly new names entering technical engagement behind them. Either way, that’s evidence the pipeline is still expanding, not just converting.

On what this looks like for revenue: initial orders and production ramp are genuinely different scales, and it helps to look at how a directly comparable company has actually described that distinction, since Sivers hasn’t broken its own numbers out this precisely yet. Aeva, on their own Q2 call, laid out almost the identical two stage structure for their NPO deal with a major hyperscaler, initial deployments starting as early as second half of 2027, then scaling into full production ramp through 2028, with the eventual run rate described as multiple millions of units annually and multiple hundreds of millions of dollars in revenue once fully ramped. That’s Aeva’s own disclosed number for Aeva’s own deal, not Sivers’, but it’s a useful template for the shape of this kind of relationship, initial orders are a much smaller qualifying batch, the real revenue shows up on the other side of that once the customer scales its own deployment.

I’d expect that H1 2027 initial order to be modest relative to what H2 2027 and beyond eventually looks like, closer to a qualification-scale batch than a steady-state run rate. The bigger revenue, the kind that actually shows up meaningfully in quarterly numbers, is more of an H2 2027 into 2028 story once Jabil and whichever of the other six follow it are both through initial orders and into ramp around the same time.

What Happens When One or Two More Convert

This is the re-rating case. Jabil already proved the mechanism works: sample laser, qualification, standard part, production order. Every additional name that follows that same path removes a chunk of execution risk from the stock, because it stops being “can Sivers actually do this” and starts being “how many more of these land, how fast.” If the tech is good, and everything from the AAOI bottleneck commentary to Sivers’ own alpha-stage engagements suggests it is, I’d expect another Jabil-style announcement from a high profile name at some point, and I think that’s the point the stock re-rates again, the same way it did across that 48 hour window in April. Not because the fundamentals changed overnight, but because the market stops discounting execution risk it no longer needs to price in.

Risk

Here’s the honest risk case. This entire near term thesis rests on Jabil actually converting from beta into production orders on schedule, and at least one or two of the other six following the same path within a similar window. If Jabil slips or doesn’t convert, and none of the six announce a Jabil-style production milestone either, the near term pluggables story breaks down.

The good news is there’s a second line of defense, XPO/NPO and full CPO are still coming behind this regardless, through GFS SCALE and Ayar Labs. But that’s a longer, later runway, and if pluggables stall out, there’s a real stretch of bag holding while the company tries to prove it can execute on that second front instead. If that also doesn’t materialize, the thesis doesn’t just weaken, it breaks, and the company’s ability to execute at all comes into question. That’s the actual downside case, not a soft one. Size any position with that in mind, not just the bull case above.

That’s the pluggables story end to end: Jabil’s actual timeline, where the other six stand, why standard products change the speed of all of this, and what I think happens when the next one converts. What stands out to me most going through all of this again is that management is playing the hand actually in front of them, leaning into pluggables because the demand and the readiness are both real right now, not chasing the CPO story before its time. I’ll keep updating this as beta turns into qual cycles and qual cycles turn into orders.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

Quick context first. Q2 for Sivers is May through July. Three months. That is not a lot of runway for a company mid transformation to show up in a top line number, and I think that gets lost when people just react to the headline.

On the two things I was actually watching this quarter, the opportunity pipeline and the Jabil timeline, they delivered. Most of what else moved the stock is either normal transition-quarter noise or stuff management already told us was coming.

That is also the whole point of doing this after every print. A company going through a transition like this one hands you new information every quarter, and that information should actually move your model, not just get filed away as confirmation of what you already believed. So after the numbers and the call itself, I want to walk through where this specific print moves my own timeline and price target.

The Numbers

• Revenue: 53.8M SEK, down 12% YoY (10% at constant FX)

• Product revenue: up 13% YoY (18% constant FX)

• Adjusted EBITDA: -35.5M SEK vs -20.9M SEK, mostly a non-cash charge tied to the share price spike, zero cash impact

• Balance sheet: 825M SEK raised plus a $12M convertible converted to equity, both closed in July, funding is no longer the constraint

• Pipeline: $1.2B as of July, up 52% since May alone

• New $4B SAM identified for optical amplifiers in optical circuit switches

Why the Stock Dropped 25%

I get why the sentiment right now feels like the sky is falling. This was a genuinely large drawdown on a stock that had already given back a lot from its highs, and that is not nothing.

I don’t think the selloff was driven by a single piece of fundamentally thesis-breaking information. Management told us months ago that 2027 is when this starts showing up in the numbers, and 2028 plus is when the long term financial model becomes visible. Q2 was never supposed to be a reflection point, and they said so again on this call. So I went back through the transcript specifically looking for what I might have missed, because a move this size usually means something.

I do not think there is a single smoking gun. A few things probably stacked together.

$SIVE is up 843% over the past year. After a run like that, a weak headline print, even one flagged well in advance, gives a lot of holders a reason to take profit or de-risk. That kind of move does not need new bad news, it just needs an excuse.

The EBITDA loss widened year over year, and even though the CFO explained the non-cash charge clearly on the call, most people skim headlines, not transcripts. A wider loss next to a revenue miss reads badly at a glance.

There is also no formal guidance here, so short term holders do not have a number to hold onto between now and Q4. When the story is entirely “trust the multi year plan,” that is a harder hold for anyone not already deep in the pipeline mechanics.

And this is a thinly traded, historically volatile stock. Moves overshoot in both directions here more than they would in a large cap.

Separately, I think there is a real argument that the market is pricing this like a public company on trailing financials, when the actual value is closer to what a private, pre-scale strategic asset in this space would command. We have seen that gap show up plenty of times across this AI infrastructure buildout. I was never trading this quarter’s print anyway. I knew going in that current earnings could not support the current market cap on their own. The entire bet is the 2027 to 2028 ramp.

Why My Focus Shifted to Pluggables

I got into $SIVE originally for the CPO (co-packaged optics) side, through the Ayar Labs relationship. I did not weight the pluggables side much at all early on.

That changed once Jabil came into the picture. Pluggables became the bridge, near term revenue while CPO matures, and Sivers is not expecting meaningful CPO revenue until 2028. Jabil buys time, and more importantly, buys cash flow while the bigger CPO story plays out.

On this call, Sivers said six more pluggable manufacturers are engaged beyond Jabil, three in alpha sample evaluation and three in technical engagement and supply assessment. That is happening against a backdrop where $AAOI and are reportedly selling everything they can produce on the pluggables side. If that demand picture is right, being a newly qualified laser supplier into a supply constrained market is a good place to be.

Q2 Watchlist: 12 of 14

Here is how I’d score every item I was tracking into this call.

Bonus items not on the original watchlist: the new $4B SAM for SOAs in optical circuit switches, and management confirming they are actively evaluating M&A.

12 of 14 confirmed or positively updated. Execution is the main risk from here, not the story itself.

What 2027 Could Look Like

The real test is Jabil. Alpha builds are done, beta builds are Q4 2026, then a qual cycle, then initial production orders in H1 2027, then a ramp in H2 2027. That is a specific, watchable timeline, and I plan to track each stage as it hits.

If Jabil converts, think about the revenue in layers: initial orders, then the ramp, then whatever comes after. The CEO has talked about multi year relationships once a pluggable customer is in production, so a converted Jabil program is not a one time bump, it is closer to an annuity.

If Jabil slips or falls through, the next step is figuring out why, then shifting attention to how the other six-plus pluggable engagements are progressing on their own. Jabil is the lead domino here, not the only one.

I think the market’s focus on the short timeline is fair. The Q4 2026 and 2027 inflection points are close now, which means management is running out of room to lean on the roadmap instead of the results. That is a real pressure point and I would not wave it away.

Here is my case for why the current price already reflects a fair amount of that risk. Even in a scenario where Jabil disappoints entirely, I think the LiDAR customer, ALL.SPACE order, and Tier 1 telecom program provide a meaningful underlying value base, with CPO and NPO (near pluggable optics, the step in between) optionality sitting behind them through GFS SCALE and Ayar Labs. Whether that translates to a $1 billion market cap is something the actual production ramps will ultimately have to prove.

On manufacturing: Sivers now has two secured foundry relationships for the indium phosphide (InP) lasers this whole story runs on, plus the Glasgow expansion underway. That is real money, probably some near term gross margin pressure while capacity comes online ahead of full utilization, and probably some dilution along the way to pay for it. But in this supply chain, if you cannot deliver volume to a hyperscaler-adjacent customer, you get designed out. Expand or die, basically. I would rather they spend the money now and have a safety net than get capacity constrained right as production orders start layering in.

On pluggables specifically, Jabil is doing more than proving out one customer, it is proving out the laser design itself. That is a big part of why I think the next six engagements do not need the typical three to five year customer acquisition to revenue runway a brand new, unproven design would need. Jabil itself was only announced this year and is already through alpha builds heading into beta. If that pace holds anywhere close for even one or two of the other six, next year’s pluggable side could have more than one program moving at once, while CPO and NPO keep maturing in the background for 2028.

Last thing here: the LiDAR customer. I think this is Aeva, based on how the CEO talked about them expanding and Sivers helping with “other business ventures” without naming them, and Aeva has separately launched its own Optical Connectivity business, which lines up with that framing. I want to do a full deep dive on timing relative to Aeva’s own hyperscaler relationship in a future issue, there is more to pull on here.

Where This Puts My Price Target

Here is the actual scenario I am running with after this call, and the number it points to.

Jabil converts and ramps on the guided timeline. Three of the other six pluggable engagements also get to initial orders in that same window, not all six, three is just a reasonable base case given how fast Jabil itself has moved.

On the CPO and NPO side, I expect real news on POET/Lumilens, GFS SCALE, and Ayar Labs late in 2027, with actual initial orders following in early 2028. That still lines up with the 2028 CPO timeline Sivers has been guiding to all along.

Getting there needs capacity, which is why the foundry buildout matters so much. In this scenario I have Sivers at roughly half of full capacity utilization by the end of 2027, then a full ramp through 2028, at which point I think they end up supply constrained and fully booked the way $LITE is right now. There is also a real M&A angle given management’s own comments on staying open to it, but I am leaving that out of this specific model to keep it conservative.

Putting a number on it: my bullish execution scenario has Sivers reaching roughly $43M USD in quarterly revenue by Q4 2027. Here is how I get there, not a forecast Sivers has given, just how I’d allocate it across what is actually converging in that window:

• Jabil pluggables ramping: ~$15M

• Other three converted pluggable customers: ~$10M combined

• LiDAR customer production: ~$5M

• ALL.SPACE SATCOM production: ~$3M

• Tier 1 telecom: ~$5M

• Remaining legacy, and other business: ~$5M

That adds up to roughly $43M. This is a bullish execution scenario, not a base case, it assumes several of these programs convert and ramp in roughly the same window. Treat it as one coherent scenario, not six independent forecasts you can mix and match.

Annualize that run rate and apply a 30x revenue multiple, and you get to a market cap of around $5.2 billion. I’m using 30x here instead of 20x because none of this $43M includes CPO or NPO revenue, which is still a 2028 plus story. The higher multiple is compensating for that optionality sitting outside the number, not for growth that’s already counted in the bridge. Against roughly 300 million shares outstanding as of the last report, that pencils out to somewhere around $17 a share, still above the old $11.25 high.

A couple of honest caveats on that number. Share count almost certainly grows between now and then given the capacity spend I just described, so the real number is probably a bit lower than $18 to $19 once that is accounted for. And the whole thing lives or dies on Jabil actually converting on schedule, plus at least a few of the other pluggables following it. I will keep updating this target every quarter as we get more to work with, that is the point of doing it this way.

From the CEO Interview

A few notes from Vathulya’s interview the day after the call.

On the Q4 inflection: he reaffirmed it directly, framing this as a pivotal transformation quarter with financing now secured and revenue programs materializing.

On the pipeline, his framing was that this is not a question of when the pipeline converts, it is converting into production orders and revenue right now, continuously, as they pull items through the funnel. My read on that: the execution risk was never really about whether Sivers can pull its own pipeline through, they have shown they can. With Jabil specifically, the open question is whether Jabil has the downstream customers lined up for Sivers’ laser. That is where my 75% conversion number comes from, and I want to be upfront that it is a subjective gut number from me, not something derived from Sivers’ own track record or anything they have disclosed. Jabil is through alpha and heading into beta, but the qualification cycle and Jabil’s own downstream demand still have to actually play out before that number means anything.

On why customer names stay private: business confidentiality, and a lot of these are high profile customers who have not authorized disclosure.

On the next 12 to 18 months, direct quote: “the strongest prospects for us to take into production are pluggables. NPO and CPO optics we see as the wave after that.” He also pointed to an existing pluggables shortage Sivers is positioned to help fill, and said more customers get named “when the time is right.”

On the capital raise, use of proceeds: fortifying the balance sheet, funding the Glasgow expansion for the hybrid manufacturing model, adding field resources to service a growing customer base, and accelerating standard product development. That last one is the piece I like most. Standard products mean a new customer can use an existing laser design instead of starting a custom build from scratch, which is exactly the mechanism behind the faster pluggables ramp case above.

On the Nasdaq listing rationale: market relevance, customer proximity, investor access, visibility, and liquidity. Good to hear him say directly that he sees it as important.

Last quote, on the technology itself: “We believe our indium phosphide manufacturing platform is world class... including our laser array technology and our optical amplifier technology.” His point was that the combination of trusted, differentiated tech plus capacity you can actually count on, through the foundry partnership, is what the market needs right now. Technology alone is not enough in a supply constrained environment, you need customers to trust you with both the design and the capacity behind it.

Reader Questions

A few questions from replies this week. I will add more as people send them.

One reader raised a few concerns together: no mention of POET or WIN Semiconductor on the call, a new foundry appearing out of nowhere (is the hybrid manufacturing pivot compensating for something), zero pre-orders on the optical side this close to year end, and a pipeline mix shift where design-in and design-win activity actually decreased while pre-design-in is what drove the 52% pipeline increase.

POET: nothing has changed on that thread specifically, this quarter’s call just did not touch it, which is not the same as a red flag.

WIN Semi: Management only talked about adding capacity, never about pulling back from an existing relationship. Nothing here reads as compensating for a problem.

No pre-orders: this makes sense given how Sivers actually sells. They are not $LITE or $AAOI, who ship a complete finished module. Sivers sells the laser into someone else’s optics, and management was clear on the call that customer qualification cycles have to clear before production orders happen. No pre-orders at this stage is normal for that kind of supplier, not a warning sign.

Pipeline mix: a one quarter dip in later stage pipeline against a jump in early stage volume is exactly what I would expect right after six new pluggable engagements got added. Those need time to mature through design-in and design-win. I would watch Q3 and Q4 before reading anything into a single quarter here.

From @mytslastory: do you think $SIVE reaches its all time high again?

Yes, I think it does, I am just not confident on timing. It happens faster if Jabil executes on the current timeline. If it does not, the path runs through the other pluggable engagements first, then CPO and NPO through Ayar Labs and GFS SCALE, which pushes the realistic timeline closer to 2028.

From @advaitavedantin: realistic pipeline conversion rate, when revenue shows up, and dilution

On conversion rates: The way they describe their own pipeline process, several rounds of discussion, Sivers gaining conviction, and the customer giving line of sight and early forecasts before anything gets dollarized, means this is already past the raw lead stage by the time it shows up in the $1.2B number. That is closer to what B2B sales would call a qualified opportunity than a cold lead, and the demand backdrop right now, with AAOI and Lumentum reportedly sold out and InP capacity itself short, is a real reason to lean toward the higher end of what is typical. A realistic range for that kind of pipeline converting into actual revenue is probably something like 50%. There is still no hard published number here though, generically or from Sivers itself, so treat that as an informed range, not a stat.

Personally, I’m hoping it lands somewhere in the 70% range given how tight the demand picture looks.

On timing, management is guiding to a Q4 2026 revenue inflection, then a broader product revenue inflection through 2027. My own expectation is early signs in Q4 or Q1, with meaningful Jabil specific revenue not landing until the second half of 2027 if it converts. I am putting that conversion at 75%, again just my own subjective estimate, not a number Sivers has published or one with any track record behind it yet.

On dilution, it is a fair ongoing concern for any company moving from R&D stage to production stage, it is part of the trade. So far I think the CEO has handled the raises and the convertible conversion thoughtfully rather than reflexively.

From @MondeloSA: Does management have the right expertise to execute this kind of shift?

I am confident here. The CEO has 25 plus years in the space, the board has been strengthened, and he correctly identified pluggables as the faster path to revenue even though CPO with Ayar Labs was the original core thesis. Bringing in the GFS SCALE relationship as a reference design win is the same instinct at work. On top of that, he has been a consistent insider buyer. That combination is what gives me confidence in execution, not just the roadmap on its own.

That is the full Q2 rundown. Story intact for me, still holding, and I will have a dedicated Aeva piece coming since I think that thread deserves its own issue rather than a few paragraphs here.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

That theme showed up loudly in mid-August, when Nvidia SVP Gilad Shainer publicly confirmed that its Spectrum-X CPO switches were in production and shipping. That mattered because for weeks, research shops had been floating the opposite story: concerns that manufacturing yields and testing challenges could push broader CPO adoption into 2028 or later. Nvidia’s confirmation, walked that rumor back fast.

Manufacturing yields and testing challenges are doing a lot of work in CPO. This piece is about the company whose entire job is that: Msscorps (6830 on the Taiwan exchange), a lab most investors have never heard of, sitting at the exact spot where “does this chip actually work” gets decided.

Every AI chip needs to talk to other chips, fast, and today that mostly happens over copper wire. Copper has a range problem. Past a certain distance or a certain speed, the signal degrades and burns a lot of power doing it. The industry’s fix is to replace some of that copper with light. Optical signals don’t carry the same distance or power penalty.

CPO is the most aggressive version of that fix. Instead of a separate optical transceiver plugged into the front of a switch, the light-generating and light-detecting components get packaged directly onto the same substrate as the chip itself. Nvidia’s Jensen Huang has been blunt about the direction of travel: use copper where you still can, but as chips scale up, scale out, and scale across, more and more of that connectivity has to happen with light instead.

The catch is that packaging light-based components this tightly is a genuinely hard manufacturing problem. That’s where Msscorps comes in.

What Msscorps actually does

Msscorps calls itself, informally, an “IC hospital.” Strip away the branding and the job is simple. When a chip doesn’t work right, or a manufacturer wants proof that a new process actually behaves the way it’s supposed to before betting a fab on it, Msscorps is one of the labs that gets called in to find out why.

That splits into two established businesses:

Materials Analysis (MA): before mass production, you need proof your materials and process steps behave the way they were designed to, at a scale invisible to the eye. Msscorps runs advanced imaging and spectroscopy (electron microscopes, ion beams, X-ray tools) to verify that.

Failure Analysis (FA): after something breaks or underperforms, Msscorps takes it apart, sometimes layer by layer, to find the exact defect. Forensics for a chip that failed a test.

Both are old, established businesses. Msscorps has been doing this since 2005, for customers that include $TSM, $NVDA, $AAPL, $AMAT, and $LRCX. Apple, Nvidia, Applied Materials, and Lam Research are named together in Taipei Times’ coverage of Msscorps’ Silicon Valley lab, while TSMC is identified as a major customer in separate reporting. Add $ASML, which Taipei Times reported adopted Msscorps’ materials analysis for its newest extreme ultraviolet lithography tools.

What’s new, and what this piece is actually about, is a third bucket: CPO inspection.

The CPO-specific piece

Msscorps has spent six years, per its own chairman, developing a specific piece of equipment: an infrared camera system that detects light leakage and defects inside a photonic waveguide, the tiny channel that carries light signals inside a CPO chip. They call it “light loss detection.”

This matters because CPO is unusually unforgiving. In a normal optical transceiver, if a component underperforms, you unplug it and swap in a new one. In CPO, the optical engine is fused onto the same package as the switch chip. If it fails after the fact, you’re not swapping a module, you’re replacing the whole switch. That serviceability problem is exactly why catching defects before a chip ships, rather than after, is worth so much.

Msscorps has priced this inspection equipment at NT$40 million to NT$100 million per unit and is targeting a commercial launch by the end of 2026, calling it a second growth engine alongside its traditional MA/FA business.

So how much of the actual CPO volume runs through Msscorps today?

Here’s the honest answer. Not much yet. CPO itself is barely out of the gate, so low current volume is exactly where a company in this spot on the timeline should be, not a red flag.

Nvidia’s mid-August announcement was specifically about Spectrum-X, a scale-out (rack-to-rack) product. The harder version, scale-up CPO (chip-to-chip, inside a single rack), is still on a 2027-2028 timeline by Lumentum and Coherent’s own recent guidance. Per a Taiwanese business paper (UDN) covering Msscorps’ CPO business, their current work there is still centered on R&D, engineering verification, and a small number of specialized test applications, not high-volume production testing.

That’s the picture right now: early, high-value qualification work with the handful of companies actually building CPO hardware, not assembly-line volume. The larger volume opportunity appears more likely to emerge over 2027-2029 as broader CPO deployments ramp, though that timeline has already slipped once and could move again.

Why this is a real bottleneck, not just a nice niche

Industry sources have been explicit that testing and yield are an actual gating factor for CPO’s timeline, not a footnote. Coverage of the Nvidia CPO announcement named silicon photonics process yield, laser coupling, and wafer-level packaging testing as the three hurdles standing between CPO and true mass production, any one of which can delay the schedule. At the OCP APAC industry summit in August, an ASE Group technical program manager said openly that the industry still lacks shared testing standards for CPO, meaning every player in the chain has to run its own independent verification process, which is slower and more expensive than it needs to be.

Zoom out further and the whole Taiwan supply chain is having a moment. AMD just committed over $10 billion to Taiwan’s advanced packaging ecosystem (ASE, SPIL, Sanmina, Wiwynn, and others), on top of Nvidia’s own build-out there. That’s not direct evidence for Msscorps specifically, it’s chip packaging capacity rather than optical testing, but it’s a sign of how much capital is flowing into the broader Taiwan AI buildout Msscorps sits inside.

That’s the bull case in one sentence: testing and yield are real constraints on an architecture companies across the AI networking ecosystem are racing to commercialize, and Msscorps is one of the few companies positioned to do the specific optical version of that testing.

The delay scare, and what it actually resolved

Worth being precise about what happened in August, because it’s easy to overstate. Before Nvidia’s announcement, research shops (SemiAnalysis among them) were floating the idea that CPO could slip to 2028 or later, and Morgan Stanley cut its 2027 optical engine shipment forecast hard, from 20-30 million units down to 6-7 million. CPO names across the board sold off on that. Nvidia’s production confirmation pushed back on the worst version of that story. It didn’t erase the uncertainty. Scale-up CPO still isn’t shipping in volume today. But it did confirm the architecture is real and moving, not vaporware.

One more nuance: that delay scare was mostly about CPO broadly, which gets lumped in with 800G/1.6T pluggable optics in a lot of coverage. $AEHR sold off hard on the same rumor even though its own exposure leans toward that broader silicon photonics ramp rather than CPO specifically. Msscorps’ upside is tied much more narrowly to CPO yields in particular, which makes it a cleaner read on this exact story, but also means it won’t automatically catch a bid on every piece of good silicon photonics news the way a broader-exposure name might.

The patent fight

One more piece worth knowing. Msscorps is currently suing a competitor, Enli Tech, for patent infringement, seeking roughly NT$200 million in damages over what it says is its core light-leakage detection patent, granted (not just pending) in Taiwan, Japan, and the US, with the US patent issued in April 2026. Enli disputes the infringement claim. It’s not public how Nvidia or other hyperscalers are actually weighing that lawsuit in their own qualification decisions. But if Msscorps’ technology remains the preferred qualified solution while the case plays out, the company’s existing lead could prove difficult for competitors to displace, regardless of how the lawsuit itself ends.

The numbers

Growth has been great and recent. Q2 2026 revenue came in at NT$635 million, up 16.58% year over year, with first-half revenue up 20.24%. Management’s own guidance, given this past December, called for NT$2.7 to 2.8 billion in full-year 2026 revenue, up from actual 2025 revenue of about NT$2.18 billion, with gross margin expected to recover from around 20% to around 30% as last year’s heavy equipment spending eases. The stock has been volatile along with the rest of the Taiwan CPO basket, up over 160% year to date but still clawing back from a rough three-month stretch tied to the broader TAIEX selloff.

Where I think this goes

Personal take, not analysis: I think Msscorps ends this decade somewhere in the $7-10 billion market cap range if the CPO thesis plays out the way I expect.

Here’s what that would actually require. Msscorps trades today around a $790-800 million market cap (converting Tuesday’s NT$25.28 billion at roughly 31.85 TWD per USD) against 2026 guidance of about $85-88 million in revenue. That works out to close to 9x forward sales.

Hold that same multiple steady, and getting to $7-10 billion means growing revenue to somewhere between $750 million and $1.1 billion, roughly 9 to 13 times where 2026 guidance sits. That’s an aggressive ask. Current growth is running 17-20% a year, and this would need something closer to 70-100%+ compounded annually for several years straight. Realistic only if CPO inspection volume genuinely scales the way the thesis assumes.

The more likely path, if it happens, probably involves the multiple expanding too, the way it already has for other names on this exact cycle. $AEHR was trading at over 60x its most recent fiscal-year revenue as of late August, even with that revenue down year over year, purely on the market pricing in future photonics test demand. Msscorps doesn’t need anywhere close to that. Even something like 20x, still a big re-rate from today’s 9x but well short of where AEHR sits, would only need revenue in the $350-500 million range to hit the same $7-10 billion mark. That’s still a large jump from $85-88 million, but a far more believable one.

Either path is a bet that CPO testing volume actually shows up at scale, not a guarantee. Treat the $7-10 billion figure as a thesis, not a target price. Both the market cap and the AEHR multiple above are snapshots, worth confirming against live quotes right before this goes out since both names move fast.

Where this fits with the rest of my coverage

If you’ve been following the $SIVE side of things, this is the other half of the same story. $SIVE and companies like it are racing to build the laser and photonic components CPO actually needs. Msscorps’ job is to make sure those components work once they’re built, and to catch the ones that don’t before they ship. Same architectural shift, different chokepoint. Shunsin (packaging), Win Semi and $TSEM (foundry), and Msscorps (yields) all sit on different links of the same chain.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

A “foundry” is a company that manufactures chips for other companies’ designs, rather than designing and selling its own. Think of it like a print shop: you bring your file, they run the press. TSMC is the famous example. X-FAB and Tower Semiconductor are both foundries, just specialists in older, sturdier chip types (analog, power, sensors) rather than the newest, smallest logic chips.

“Silicon photonics” means using light instead of electricity to move data around inside a data center. Electrical wires lose energy as heat and slow down over distance; light through fiber doesn’t have that problem nearly as badly. The catch: silicon is good at switching electricity but bad at making light. You need a different material, usually indium phosphide (InP), to actually generate the light, then you have to combine it with silicon to build the finished chip. That combining step is the hard, expensive part everyone in this story is trying to solve.

“CPO” (co-packaged optics) is the next step in that idea. Instead of the light-making chip sitting off to the side connected by wires, you place it right next to the computing chip, in the same package. Less distance for the electrical signal to travel before it becomes light means less wasted power and higher speed, which matters enormously once you’re wiring together 100,000 AI chips in one building.

“800V” and SiC/GaN are about power, not data. AI data centers and electric vehicles both need to move huge amounts of electricity efficiently. Moving to higher-voltage systems (800 volts instead of the older 400V standard) loses less energy over the same wiring. Silicon carbide (SiC) and gallium nitride (GaN) are the special materials used to build the switches and converters that make 800V systems work; regular silicon can’t handle it as efficiently.

Why governments are paying for this: the companies that dominate chip manufacturing today are mostly in Asia and the US. Europe has decided that’s a strategic risk, and is spending real public money (the EU “Chips Act”) to build manufacturing capability at home, including specifically in photonics. That’s not marketing language. It’s actual line items in actual EU budgets, and it’s the reason a small Belgian-German company keeps showing up next to Nvidia’s name.

Keep these five ideas in your pocket. Everything below is just these five things, applied to one specific, small, unglamorous-looking European company.

Part 1: The Precedent, What Tower Semiconductor Already Proved

Before X-FAB, there’s Tower Semiconductor, because Tower’s story is the whole reason this thesis exists.

In February 2022, Intel agreed to buy Tower, a specialty foundry, for about $5.4 billion. The deal died in August 2023 after 18 months stuck in Chinese regulatory review: Beijing simply never signed off. Tower’s stock, which had been trading below the deal price for a while (the market didn’t believe it would close), fell further. At the time, Tower also had a genuinely bad quarter: revenue down 16% year-over-year. By any normal read, this was a washed-up company that had just failed to get bought.

Here’s what actually happened next: from that August 2023 low (around $34/share) to August 2026, Tower’s stock is up roughly 650%. Market cap is now north of $28 billion, more than five times what Intel offered for the entire company. Bank of America just put a $367 price target on it. Tower just reported record quarterly revenue, up 24% year-over-year, and raised its own 2028 target to $3.6 billion in revenue.

Why did this happen? Not because Tower suddenly invented something new. Tower had been quietly building a silicon photonics manufacturing platform (called PH18) for years before anyone cared. Coherent, a real, product-shipping optics company, had already publicly named Tower as a key supplier back in March 2024, over a year before the stock did anything dramatic. What changed wasn’t the technology. What changed was that AI data centers made “who manufactures the optics” suddenly one of the most important questions in tech, and the market went looking for an answer it had been ignoring.

The pattern, stated plainly: a boring specialty foundry, written off for reasons that had nothing to do with its actual capability, sitting on a real photonics platform nobody was pricing, until, all at once, everybody was.

That’s the pattern this whole thesis is betting X-FAB might repeat. Now let’s meet the company.

Part 2: Meet X-FAB

X-FAB Silicon Foundries is a Belgian-headquartered (the actual operational hub is Erfurt, Germany) specialty foundry focused on Europe. It doesn’t design chips, it manufactures other companies’ designs, the same business model as Tower. It runs six wafer fabs across Germany, France, Malaysia, and the US.

It has three business units:

• Smart CMOS & SOI: the historical core, analog and mixed-signal chips (think sensors, power management)

• Wide Bandgap: silicon carbide and gallium nitride power chips, the ones relevant to that 800V story above

• Microsystems & Photonics: MEMS sensors, plus the new silicon photonics platform

Its customers need chips that last 15 to 20 years reliably (cars, medical devices) rather than the newest, smallest transistor, so it deliberately doesn’t compete with cutting-edge foundries like TSMC. That’s a real, defensible niche: once a chip is qualified on an X-FAB line, switching foundries means re-qualifying an entire automotive or medical supply chain, which is slow and expensive enough that customers routinely pay in advance just to hold their capacity slot.

Part 3: The Three-Legged Thesis

Leg 1: SiC/GaN and the 800V Shift

X-FAB already has a real, growing, currently-generating-revenue business in silicon carbide and gallium nitride power chips. This segment grew 152% year-over-year recently, off a small base. As AI data centers and EVs shift to 800V power systems, demand for these chips is expected to keep climbing. X-FAB’s Lubbock, Texas facility is, per the US Department of Commerce, the only high-volume SiC foundry in the United States, which is also why it signed preliminary terms for up to $50 million in proposed US CHIPS Act funding. On the company’s own recent earnings call, management said the 800V transition “is expected to support demand for both SiC and GaN” for X-FAB, and disclosed several new SiC design wins tied to data-center customers. This leg is real, already happening, and doesn’t depend on anything unproven working out.

Leg 2: Silicon Photonics and CPO

This is the exciting, unproven leg, and the one that determines whether X-FAB is “a solid power-semiconductor company” or “the next Tower.”

X-FAB is building a silicon photonics platform using a technique called micro-transfer printing (MTP): essentially, precisely stamping thousands of tiny light-making chiplets (made of indium phosphide, gallium arsenide, or other materials) onto a silicon base, rather than trying to grow incompatible materials together directly. It’s also working with thin-film lithium niobate (TFLN), a material good at extremely fast optical modulation, through a partnership with Swiss-Belgian firm LIGENTEC. The laser chiplets themselves are expected to come from a partnership with SMART Photonics, a Dutch InP specialist.

The centerpiece is photonixFAB, a €48 million EU-funded consortium that X-FAB leads. This isn’t just a funding line; real, credible names are involved as participants, including Nvidia, Nokia, and imec (the respected Belgian chip research institute), alongside Luceda Photonics (design software) and SMART Photonics. The stated goal is establishing a European silicon photonics value chain with a real path to high-volume manufacturing, spanning several different demonstrator applications from data-center transceivers and optical switches to sensing.

That’s a genuinely strong signal. You don’t often find Nvidia’s name attached to a pre-commercial European foundry project. But be precise about what it is: this is Nvidia and Nokia participating in and evaluating the technology, not a signed commercial supply agreement. If X-FAB’s process doesn’t hit the yields and reliability these partners need, there’s no contract obligating anyone to use it afterward. It’s a real vote of technical seriousness, not a purchase order.

On a recent earnings call, X-FAB’s own language was: “we continue to expect photonics volume production to start in 2028.” The company has pointed to multiple photonics opportunities in development with industry partners. The 2028 date itself is a useful, specific anchor: not “someday,” a stated year, from the company itself.

Leg 3: European Sovereignty Money

This leg doesn’t depend on anyone liking the technology at all. It’s about governments wanting a European alternative to exist, period.

• Fab4Micro, the new cleanroom being built at X-FAB’s Erfurt site, received a €127.4 million grant (roughly €80M German federal, €47M from the state of Thuringia), approved by the European Commission in December 2025 under the original EU Chips Act. That’s real money already committed, not a promise. Thuringia’s head of government called it “a key project for the technological sovereignty of Germany and Europe.” Fab4Micro’s own scheduled production start is end of 2028.

• Chips Act 2.0: a new, broader EU proposal, presented June 3, 2026 as part of the EU’s “Technology Sovereignty Package,” explicitly names photonic integrated circuits as a priority area, alongside quantum chips and advanced packaging. Important caveat: this is still a proposal working through the EU legislative process (Commission, then Parliament, then Council, then final adoption): real, and a genuine tailwind if it lands, but not yet enacted law. Don’t treat it as money in the bank.

Put together: even if the photonixFAB bet doesn’t pay off commercially, X-FAB is getting non-dilutive government capital to build capacity anyway, for reasons that have nothing to do with whether the technology succeeds commercially. That’s a real cushion under Leg 2’s uncertainty, not a guarantee, but a genuine one.

Part 4: What Just Got Confirmed Recently

A few specific, dated things happened close together that are worth knowing as a cluster, because they reinforce each other:

• June 23, 2026: The Fab4Micro €127.4M grant was formally presented in Erfurt, the same day Bernstein upgraded X-FAB from Market Perform to Outperform, more than doubling its price target from €5.00 to €12.80.

• Management confirmed direct exposure to both CPO and the 800V transition explicitly. That’s useful, because some skeptics had specifically argued X-FAB had no real exposure to either theme. That specific bear argument no longer holds up as stated.

None of this changes the fundamental timeline: production is still 2027 to 2028 territory.

Part 5: The Full Timeline

The single most important thing to internalize from this table: “ready” (2027) and “volume” (2028) are two different milestones, a full year apart. They aren’t the same thing, and conflating them is the easiest way to over-promise this stock to yourself.

Part 6: The Skeptical Case

Scale gap is real and large, and getting larger. Tower’s silicon photonics business alone hit a $680 million annualized run rate in Q2 2026, up 270% year-over-year, and management is targeting $1 billion by Q4 2026. Tower has also disclosed roughly $1.3 billion in signed customer contracts for 2027 silicon photonics revenue, backed by $290 million in customer prepayments already received to secure that capacity. That’s not a platform waiting for validation, that’s contracted and prepaid demand, even without Tower having disclosed exactly which customers are behind those specific contracts. X-FAB’s entire Microsystems & Photonics segment (which blends in non-photonics MEMS too) ran at roughly $135 million annualized in its best quarter. Tower, GlobalFoundries, and SilTerra all have working commercial relationships with real optics companies today; X-FAB doesn’t have a named, shipping commercial customer yet. photonixFAB is evaluation-stage, not revenue.

The core business is genuinely struggling right now. Automotive, X-FAB’s largest market, was down 19% year-over-year last quarter. The company posted a net loss. Capacity utilization sits around 60%. The photonics story is exciting, but it’s layered on top of a business currently in a real cyclical trough, and even management won’t commit to exactly when that trough ends.

Management’s targets have a track record of moving. A 2023 target of $1.5B revenue by 2026 got pushed to 2030, then softened into vague language with no specific number attached at all. Any forward-looking number X-FAB states, including the CEO’s own mention of “$300 million” in long-term data-center revenue potential, should be read as direction, not commitment.

Part 7: The Actual Playbook, What to Watch and What Each Outcome Means

Here’s what to actually track, and what it tells you if it happens:

Two dates worth circling now: September 1, 2026 (investor update) and October 29, 2026 (Q3 earnings), the next two moments this thesis gets tested.

My 2030 Prediction

Everything above is analysis. This part is a personal call, and I want to flag it clearly as possible.

Tower, I think, keeps compounding past its own 2028 target. The company is already guiding to $3.6 billion revenue and $1.2 billion net profit by 2028, up from where it stood just a couple of years ago, and CPO demand isn’t a 2028 story that stops, it’s the early innings of AI data centers rebuilding their entire interconnect layer around light instead of copper. If that keeps scaling into 2029 and 2030 the way it has since 2023, I’d expect Tower to be trading north of $75 billion in market cap by the end of the decade. That’s roughly 2.6x from today’s ~$28 billion, which is actually a smaller multiple than the roughly 5x move it already made off its 2023 low in just three years.

X-FAB, if it actually executes on all three legs, I think lands around $10 to 15 billion by 2030. Off where it’s traded through most of 2026 (roughly $1.2 to 1.7 billion depending on the day), that’s somewhere between 6x and 12x. That sounds aggressive until you compare it to what Tower already did on a similar timeline. The difference is that X-FAB’s move requires more than growth, it requires the market to re-rate it from “depressed automotive supplier” to “credible photonics foundry” at the same time, which is a bigger lift than Tower’s move, since Tower was already being taken seriously as a foundry the whole way through.

What has to be true for the X-FAB number specifically: the SiC/GaN business keeps compounding at anything close to its current growth rate, the photonics platform lands at least one named commercial customer once 2027 ramp-up readiness arrives (not just Nvidia evaluating it), Fab4Micro’s 2028 volume production actually ships on schedule, and Chips Act 2.0 comes through with real capital rather than staying a proposal. Miss two or three of those and $10 to 15 billion by 2030 doesn’t happen, this becomes a smaller, still-respectable re-rating instead.

X-FAB is not, today, the next Tower Semiconductor. It’s a smaller, earlier-stage, currently-struggling specialty foundry with a real (if unproven) photonics platform, a genuinely strong technical relationship with Nvidia and Nokia that stops short of a commercial contract, and multiple layers of non-dilutive government money that reduce, but don’t eliminate, the risk of getting there.

What changed recently isn’t the fundamental timeline. It’s that the market’s read on this has visibly started to split: two credible analyst upgrades, explicit management confirmation of the CPO/800V exposure bears had denied, and a second wave of EU sovereignty funding all landing in the same few months. That’s not proof the thesis is right. It’s evidence that more people with real analytical resources are starting to ask the same question this piece is asking, instead of dismissing it outright.

The honest framing, one more time: Tower shows what the market eventually pays for a specialty foundry’s photonics optionality once it notices. X-FAB may be several years earlier in that same process, which is exactly why the potential upside is large.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

What would Sivers actually have to become for an $80B+ valuation to make sense? Not “can the stock go up a lot,” every small-cap story can theoretically do that, but specifically: what combination of revenue scale, margin, and market multiple gets you there, and how much of the path to that combination has actually happened versus still needs to.

The starting point

Where Sivers is today:

• Opportunity pipeline: $799M in Q1 2026, up 77%. Pipeline isn’t revenue, it’s a funnel that still has to convert through qualification and production.

• Core technology: InP laser technology and DFB laser arrays, the same laser category behind the CW (continuous-wave) laser scarcity story showing up across this whole sector.

• Manufacturing model: fab-light, scaling through foundry partnerships rather than owned fab capex.

• Financing: a SEK 700M directed share issue completed in July 2026 (approximately $70M at the time, 12,280,701 new shares), several times oversubscribed, with approximately 3.3% dilution on a fully diluted basis. A $6.6M second-year U.S. CHIPS Act/Microelectronics Commons award for the EW STAR program, officially announced May 2026.

• Existing production business: Sivers already has production activity outside the AI-photonics story. Its lead automotive LiDAR customer is expected to ramp production from Q4 2026, with Sivers previously estimating $28M-$53M of cumulative revenue from that customer over 2026-2030.

• U.S. listing: evaluating a potential Nasdaq dual listing since April 2026, with continued preparation since, including upgrading financial reporting to PCAOB standards. A July 2026 update reiterated intent to complete the listing process over the next few quarters. Actively progressing, not yet completed.

The roadmap, in four phases

Phase 1: Build the laser platform (mostly complete, scaling continues) The core InP/DFB laser technology has been demonstrated and is shipping. What’s still moving is scale: manufacturing partnerships, customer sampling, and the pipeline growth cited above.

Phase 2: Prove downstream relevance (underway) Ayar Labs, Jabil, GFS reference laser, POET, and others.

Phase 3: Convert development into production (required, with meaningful AI-photonics volume still ahead) This is the actual bottleneck, and probably the single most important idea in this piece. The path from here is sample, qualification, design win, production order, volume production, recurring revenue. For the downstream AI-photonics relationships above, nothing has yet been demonstrated at meaningful volume-production scale.

Phase 4: Move further downstream (required, not yet started) This is where the acquisition thesis comes in. Each step further downstream, laser to ELS to optical engine to transceiver to full module to system, potentially captures more revenue per unit, more gross profit dollars, more direct customer ownership, and a richer valuation multiple. Lumentum’s own history is the reference case for this logic. No acquisition has been announced. New board members with M&A backgrounds are a signal of intent, not evidence of execution.

The Lumentum Precedent

Lumentum traded around $2.7-3.0B in market cap around the time of its Cloud Light acquisition in November 2023, a deal reported at approximately $750M. Cloud Light was generating more than $200M of LTM revenue at the time. Lumentum said the acquisition would more than double its cloud intra-data-center infrastructure revenue within 12 months and expand its served market more than fivefold. Lumentum’s market capitalization is around $60B+ today. The roughly 20x.

Sivers Scorecard

Completed

• [x] SEK 700M capital raise, oversubscribed

• [x] Board additions with M&A backgrounds

• [x] $6.6M second-year CHIPS Act award (EW STAR program)

• [x] Partnerships established:

• [x] Pipeline expansion to $799M (Q1 2026), up 77%

Underway

• PCAOB-standard reporting preparation

• Manufacturing and foundry capacity scale

• Nasdaq dual listing

Still Required

• Customer qualification converting to design wins

• Volume production

• Major hyperscaler-related production orders

• A meaningful downstream acquisition

• Billion-dollar-plus revenue

• Sustained profitability at that scale

• A market re-rating to a platform, rather than components

What kills the thesis

• No meaningful move downstream, or a bad acquisition. The $80B thesis needs Sivers to capture substantially more of the optical value chain, whether through M&A or an equivalent organic move. A bad acquisition could destroy value rather than create it, and either way, the reverse-engineered math above shows revenue scale and market multiple are their own independent hurdles even once a downstream move happens.

• Underway doesn’t convert to required. “In development” is a different outcome than “in volume production,” and plenty of development-stage optical programs never make that jump.

• Dilution outruns value creation. More capital raises are likely on this path; if they’re not matched by proportional value creation, existing holders don’t participate in the upside the scorecard implies.

• CPO/NPO adoption is slower than the cycle currently assumes. Phase 3 and 4 both lean on the architectural shift toward CPO continuing to accelerate on roughly its current timeline.

• The Lumentum comparison turns out to be more cycle-driven than company-driven. If the AI optical upcycle cools, the multiple expansion Sivers needs may not be available at any revenue scale, regardless of execution.

Bottom line

Some of this roadmap is completed and dated: the raise, the board hires, the CHIPS award, the listing preparation, the partnerships getting established. Some of it is genuinely underway: the pipeline growth is a real number, partnerships are real. But the items that would actually justify $80B, qualification converting to volume, the acquisition, the jump from tens of millions to billions in revenue, are still required, not underway.

The $80B outcome isn’t a prediction. It’s what happens if every required box above gets checked, at a scale several steps beyond where the company sits today, while a favorable AI optics cycle keeps cooperating in the background. Right now the scorecard has real green checks on it. It also has a long required column that hasn’t moved yet, and that’s the honest state of the thesis, trackable, quarter by quarter, rather than settled.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

The management quotes below are drawn from AAOI’s most recent earnings call, as relayed in investor commentary. The specific quarter isn’t pinned down here, referenced instead as “AAOI’s latest call” throughout. The $471M/month figure and the mid-2027 timeline have been confirmed against the transcript.

What has to be true

An $80B+ market cap for AAOI by 2027 isn’t one bet. It’s three, stacked on top of each other, and all three have to land:

1. The company has to actually deliver $471M/month in transceiver revenue. That’s a demand-and-capacity execution question.

2. It has to happen on roughly the timeline management pointed to: mid-2027. A year late and the whole scenario shifts.

3. The market has to be willing to pay somewhere around 14-15x sales for that run-rate. That’s not a business question at all. It’s a sentiment question, and it’s the leg with the least company-specific control.

Miss any one of the three and $80B doesn’t happen, even if the other two hold exactly as described.

Leg one: can the company actually make the number?

The setup

The last two years of AI infrastructure buildout turned optical transceivers from a sleepy networking component into one of the more contested chokepoints in the entire AI datacenter stack. Modern AI GPU clusters require enormous amounts of east-west and scale-out bandwidth, and that bandwidth increasingly moves over optics rather than copper as speeds climb past 400G into 800G and now 1.6T.

Two things are happening at once. Pluggable transceiver demand went vertical, hyperscalers need more 800G and 1.6T modules than the industry has historically been able to produce, because cluster sizes and interconnect density both scaled faster than the supply chain did. And co-packaged optics (CPO) emerged as the next architectural shift: instead of a pluggable module sitting at the edge of a switch faceplate, CPO places optical engines much closer to the switch ASIC, cutting power consumption and improving bandwidth density, at the cost of a much harder integration problem and much higher switching costs once a platform commits to a design. Implementations vary across vendors, the broad tradeoff holds either way.

AAOI sits at the intersection of both trends. It’s a long-time maker of pluggable transceivers and, critically, the indium phosphide (InP) laser chips that go inside them: the same laser technology CPO architectures also need as a light source. That overlap is the entire subject of this piece.

How AAOI got here

AAOI’s history explains the volatility premium the market has historically applied to this stock: an early boom-bust cycle tied to customer concentration is part of why longtime holders stay disciplined about that risk specifically.

The 2023-2025 stretch looks like a second act. The AI datacenter buildout reignited demand for exactly the high-speed optical products AAOI has spent years manufacturing, and the current capacity buildout is centered on expanding its Houston-area manufacturing footprint. AAOI announced in April 2026 that two new buildings at its Pearland campus, roughly 388,000 square feet combined, would bring the Houston-area footprint toward 900,000 square feet, with a stated goal of supporting up to 700,000 units/month of 800G and 1.6T production in Houston and roughly 350% additional laser fabrication capacity by the end of 2027. A July 14 follow-up announcement confirmed construction had actually begun on the two properties. Sugar Land remains the company’s historical base; Pearland is the site of this specific expansion. Capex figures behind the buildout aren’t detailed here, worth pulling from the transcript or an investor presentation if the dollar cost matters to the piece. That buildout is the likely backdrop for everything management said on the call this piece is built around, though calling it “betting on capacity ahead of demand” is this piece’s read of the sequence, not something management said in those words: the company built out capacity, demand showed up faster than even the expanded capacity, and now the company is rationing laser output toward its own highest-priority business.

That’s what makes the CPO comments below matter. This isn’t a company saying it doesn’t want the CPO business. It’s a company saying it’s sold out of the input both businesses need.

What management actually said

Three takeaways stand out from the call material:

AAOI is effectively out of first-generation CPO laser deployments, not by choice, but by capacity. Management’s framing was direct: “We just can’t make enough of them [CPO lasers] to be involved in their current first-generation deployments because there’s just not enough capacity. We have to prioritize our ability to make lasers for our own transceivers first.” Read literally, that’s a company turning away business. Read in context, it’s a company confirming its core transceiver demand absorbs every unit of laser capacity it can build, with nothing left for merchant CPO laser sales. If AAOI is capacity-constrained out of the CPO laser market, that shrinks the pool of qualified independent CPO laser suppliers for everyone else building CPO systems, and the list of qualified Western suppliers who can serve that need isn’t long.

The transceiver business itself is capacity-constrained, not demand-constrained. Management’s language: “The customer demand is 20%-40% higher” than expanded built-out capacity, arriving “from several big customers almost every week.” Lasers were named specifically as “the biggest bottleneck right now for the transceiver business.” Laser fabrication, particularly InP-based DFB and EML lasers used in high-speed datacom optics, is a genuinely hard manufacturing problem: low wafer yields, long qualification cycles, a narrower base of equipment and process expertise than silicon manufacturing. Adding capacity here isn’t “build another line.” It’s standing up a process technology with a multi-quarter yield learning curve.

China is described as multiple years behind on CPO-grade DWDM lasers. Management characterized China as “easily at least two, three years or even longer” from having CPO-grade DWDM laser capability. That tracks directionally with how the InP laser supply chain has generally been described elsewhere, but it’s a single earnings-call soundbite from a company with obvious incentive to frame its competitive position favorably. Treat it as management’s stated view, not an independently verified industry assessment.

Why lasers specifically are the chokepoint

A laser chip is a tiny component relative to a full transceiver module, so it’s worth being precise about why it’s the constraint rather than assembly, testing, or packaging capacity. High-speed datacom transceivers use InP-based laser diodes, commonly DFB or EML designs, as the light source. Unlike the silicon photonics and passive components that make up much of the rest of a transceiver’s bill of materials, InP laser fabrication runs on a narrower base of specialized equipment, has historically had lower yields, requires long burn-in and reliability testing before qualification, and draws from a much smaller global base of experienced suppliers. That combination is exactly why “add capacity” for lasers is a multi-quarter-to-multi-year undertaking, and why a company that controls its own laser supply, as AAOI does, has a structurally different cost and availability position than a transceiver maker buying lasers from a merchant supplier.

Bottom line on leg one: the constraint management describes is capacity, not demand. That’s a fundamentally different risk profile than most growth stories, where the market has to be convinced demand shows up at all. Here, the open question is whether the Pearland buildout, and whatever additional capacity comes with it, can move fast enough to catch the demand that’s already there.

Leg two: does 2027 hold up?

Management pointed to mid-2027 for reaching $471M/month, specific enough to function as a real checkpoint. The figure breaks down by product generation: roughly $90M/month from legacy 100G and 400G, $217M/month from 800G, and $164M/month from 1.6T, a mix worth watching evolve as much as the total. If capacity, not demand, is genuinely the constraint, the path to that run-rate is a checklist of execution items rather than a question of whether customers show up:

Laser capacity.

• Where it stands: demand running 20-40% above built-out capacity; the Pearland buildout targets roughly 350% additional laser fabrication capacity by the end of 2027.

• What has to happen by 2027: AAOI has guided to over 930,000 monthly 800G/1.6T units by the end of 2027, with more than half of that production coming from Texas. That’s the number to hold the company to; worth noting it’s an end-2027 capacity figure, while the $471M/month revenue target is dated to mid-2027, two checkpoints on the same buildout, not the same moment, so they shouldn’t be divided into each other for an implied price point.

Customer concentration.

• Where it stands: several large customers, cadence “almost every week.”

• What has to happen by 2027: broader qualification across more hyperscaler programs, reducing lumpiness.

Product mix.

• Where it stands: 800G is the largest slice of the mid-2027 target at $217M/month, with 1.6T at $164M/month and legacy 100G/400G at $90M/month.

• What has to happen by 2027: the 1.6T share keeps climbing, since it carries the highest dollars-per-unit of the three.

CPO optionality.

• Where it stands: not a near-term revenue driver, capacity fully absorbed by core demand.

• What has to happen by 2027: stays a longer-dated call option, not part of the 2027 number.

China competition.

• Where it stands: multi-year behind on CPO-grade lasers.

• What has to happen by 2027: stays a defensibility window, not a permanent moat.

The single most important variable here isn’t any one demand data point, it’s management’s credibility on delivery dates. Laser fabrication yield ramps have a well-earned industry reputation for taking longer than guided, and this entire leg lives or dies on execution timelines the company controls but hasn’t yet proven out at this scale.

Bottom line on leg two: treat 2027 as a real target worth tracking against, not a locked-in outcome. If the date slips a year, the $80B case slips with it, even if every other assumption in this piece holds.

Leg three: will the market pay 14-15x sales for it?

This is the leg that has nothing to do with AAOI’s own execution. A company can hit every operational target on this list and still not be worth $80B, if the market simply isn’t willing to pay a rich multiple for it.

The math first. $471M/month annualizes to $5.65B. To clear an $80B market cap off that revenue base requires a multiple of roughly 14.2x sales. This piece uses a straightforward market cap ÷ sales approach rather than enterprise value ÷ sales, which skips the net cash/debt adjustment an EV multiple would require and gets straight to equity value:

What has to be true for the market to pay it. A double-digit sales multiple isn’t unprecedented for this sector during an AI-infrastructure upcycle, richly-valued optical and photonics names have traded there before when the market believed the growth was durable. But it requires the market to buy several things simultaneously: that the growth is multi-year rather than a one-time pop tied to a single product cycle, that gross margins expand as the business scales rather than staying compressed by ramp costs, that the InP laser chokepoint functions as a real moat rather than a temporary bottleneck competitors eventually route around, and that the broader AI capex cycle keeps expanding rather than decelerating. Every one of those is a belief, not a fact, and beliefs like these are exactly what re-rate fastest in both directions.

Bottom line on leg three: this is the fragile leg. Operational execution can be tracked quarter by quarter. Market sentiment about a multiple can compress in a single earnings reaction, regardless of how well the other two legs are executing.

What breaks this

Each leg has its own failure mode, and a real bear case only needs one of them to go wrong.

Leg one risk, execution and concentration. The thesis assumes AAOI can actually build out the Pearland expansion, and whatever InP laser capacity comes with it, in time for a 2027 run-rate. Customer concentration isn’t just an old AAOI problem, either: Microsoft accounted for 28.8% of 2025 revenue, down from 43.7% in 2024 and 46.6% in 2023, and the top three customers together made up roughly 95% of Q1 2026 revenue. The dominant name changed since the Amazon era, but the concentration itself didn’t go away, it’s just playing out against a much bigger opportunity now, which cuts both ways: more upside if it holds, more at stake if any one of those three shifts its order pattern.

Leg two risk, timeline slippage. Laser fabrication yield ramps have a well-earned reputation for taking longer than guided. A 2027 target that slips to 2028 doesn’t just delay the story, it compresses the multiple the market is willing to pay in the meantime, since a further-out payoff gets discounted harder.

Leg three risk, multiple compression. AI-infrastructure-adjacent names have re-rated sharply as a group. A broader deceleration in hyperscaler capex growth, or a rotation away from the theme, could compress multiples across the whole comp set regardless of AAOI’s own execution. Rich multiples are also simply more fragile than cheap ones, there’s more room to fall.

Cross-cutting risks. Scaling laser fab capacity is capital-intensive, and growth phases at this company have historically coincided with equity issuance, worth checking the current share count trend, not just the current count, before finalizing any per-share math. CPO itself could become a medium-term disintermediation risk to AAOI’s core pluggable business, not just a near-term opportunity it’s missing out on, if CPO architectures scale faster than expected, pluggables could face real substitution pressure over the back half of this decade. And the entire capacity-constraint narrative in this piece traces back to one earnings call; cross-referencing against the actual transcript, subsequent commentary, and independent industry data would meaningfully strengthen it before publication.

Bottom line

The signal from the call, taken at face value, is unusual in a good way: a company describing its bottleneck as “we can’t make enough” rather than “we’re not sure anyone wants this.” That takes leg one further than most growth stories ever get. Leg two is a real, if imprecise, commitment, 2027 is specific enough to hold management to. Leg three is where the case gets genuinely speculative, since it depends on the market’s mood about a multiple more than on anything AAOI itself controls.

All three legs holding at once is what an $80B+ AAOI actually requires. Any one of them breaking is what a much smaller AAOI looks like instead.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

The unexpected started with Aeva’s Q2 earnings call last week on Aug 5th, 2026. Aeva announced that it was entering the optical connectivity market for AI data centers, signed a joint development agreement, and said that their product would ultimately be deployed by an unnamed major hyperscaler. That immediately raised a question for me: who is supplying the laser? I think it’s Sivers.

Given Sivers’ announcement about an unnamed strategic LiDAR customer, the timing of its Q4 2026 production ramp, and the fact that Sivers already makes the type of high-power CW DFB lasers used in FMCW LiDAR and optical engines, I think there’s a real, worth-investigating connection between these two companies.

On March 11, 2026, Sivers announced a “strategic LiDAR customer” had integrated its continuous wave DFB lasers and optical amplifiers into its platforms. Production is expected to ramp in Q4 2026, with Sivers estimating $53 million to $138 million of revenue over the customer’s product lifecycle.

Then on August 5, 2026, Aeva launched a new Optical Connectivity business focused on AI data centers, and confirmed a signed joint development agreement with an unnamed optical engine provider. That partner will integrate Aeva’s optical source into a module for an unnamed major hyperscaler. Deployment is targeted for the second half of 2027, with production ramping in 2028. Everything unnamed.

Does Sivers make the kind of laser Aeva needs?

Aeva’s FMCW LiDAR needs a laser with narrow linewidth, stable single-mode operation and the ability to generate the frequency chirp required for FMCW sensing. Automotive systems also generally use the 1550nm range because of the eye-safety advantages.

Sivers has a product line specifically for FMCW LiDAR. Its own specs describe high-power ridge-waveguide DFB lasers, narrow linewidth, wavelength stability, single-mode operation, automotive reliability and up to 100mW output. Sivers’ broader InP platform covers 1270nm to 1650nm. So this isn’t just a generic Sivers laser that happens to fit the specs. Sivers is specifically marketing these lasers for FMCW LiDAR.

Would Aeva reuse an existing supplier?

Aeva’s own CTO said the new Optical Connectivity business would leverage its foundation “along with Aeva’s existing high-volume manufacturing and foundry supply chain.” That’s probably the most interesting piece of wording from the announcement. Aeva isn’t describing this as a completely new supply chain built from scratch. It’s building the new business around infrastructure and relationships it already has.

If Sivers is already supplying the lasers for Aeva’s LiDAR platform, then using that same supplier for the new business would make a lot of sense. The supplier would already be familiar with Aeva’s requirements, and Aeva wouldn’t have to start from zero with a new laser relationship. Again, that’s assuming Sivers is actually the existing supplier.

What about Lumentum and Coherent?

I wanted to look seriously at the alternatives because otherwise it’s easy to make the Sivers theory fit whatever you want it to fit. The two obvious alternatives are Lumentum and Coherent. Both are much larger than Sivers and have established businesses around lasers and optical components. Lumentum in particular markets ultra-high-power lasers for CPO.

If the question is simply who has the strongest credentials in high-power optical components, Lumentum is probably ahead of Sivers. But there is a capacity question. NVIDIA has committed $2 billion directly to Lumentum, along with major purchase commitments and future capacity access for advanced laser components. There has also been analyst commentary around major Western suppliers being heavily booked through 2027 and into 2028.

That doesn’t mean Lumentum can’t be Aeva’s supplier. It just makes availability worth considering. If Aeva needs a supplier for a program moving toward deployment in 2027 and production in 2028, an existing supplier with available capacity could be more useful than simply going with the biggest name in the industry. Coherent has a similar issue.

So when I compare the three, Lumentum and Coherent are clearly the bigger and more established names. Sivers’ argument is different. It’s technically capable, it may already have an existing relationship with Aeva, and it could potentially have capacity available when some of the larger suppliers are already heavily committed.

Who is the signed hyperscaler?

On the earnings call, Aeva’s CEO mentioned hyperscalers such as Amazon, Microsoft and Google.

Of those, AWS is the one I'd lean toward right now. The reason isn't that I found anything tying AWS directly to this deal. I haven't. It's because AWS has already announced a collaboration with NVIDIA to integrate NVLink Fusion into its next-generation Trainium4 infrastructure. That puts AWS directly in the ecosystem Aeva is now targeting with its Optical Connectivity business. Google and Microsoft are both very possible as well.

If I had to pick today, I’d put AWS first, followed by Google and Microsoft.

This is my own opinion, not financial advice. I am not a financial advisor. Do your own research before making any decisions.

I’ve opened a dedicated account holding only AAOI and SIVE, 500 shares of AAOI and 5,000 shares of $SIVE, and I’m going to track every dollar of it publicly over the next years.

My goal is simple: build this account into more than $1 million by riding what I believe will be one of the biggest infrastructure buildouts of the AI era.

Will it happen? I don’t know. Maybe the thesis falls apart next year. Maybe I sell early because the facts change. Or maybe this ends up being one of the best investment decisions I’ve ever made, or one of my most expensive mistakes. Either way, by 2030 we’ll know the answer, and everyone following along will see every major decision, every update, and every mistake along the way.

That’s exactly why I separated this from my main portfolio. My primary account remains diversified and includes a number of other positions, including my 49,500 shares of $SIVE. This second account exists for one reason: to publicly test one concentrated AI infrastructure thesis built around two companies that I believe are positioned to benefit from one of the biggest technology shifts of the next decade.

This isn’t financial advice, and it isn’t how I’d recommend most people invest. It’s a high-risk, high-conviction experiment that I think is worth documenting in real time.

If I’m wrong, everyone will see it. If I’m right, you’ll have watched it happen in real time.

Here’s why these are the only two companies in the account.

Why AAOI gets the bigger weighting right now

The short version: AAOI already has real, disclosed orders landing today, while Sivers’ biggest opportunities are still earlier-stage. That difference in timing is the whole reason for the weighting.

The orders are real and they’re accelerating. AAOI landed its first-ever 1.6T transceiver order in March 2026, over $200 million from a single hyperscale customer. Around the same time, it picked up $124 million in separate 800G orders, an initial $53 million order followed by a $71 million upsize. Full-year 2026 revenue guidance has been raised to over $1.1 billion, more than double what the company did in 2025. This isn’t a story about what might happen. Orders are landing now. For me, that’s what separates AAOI from many AI stories. The demand isn’t theoretical anymore, it’s already showing up in signed purchase orders and guidance.

Pluggables aren’t going anywhere, and that’s the point. AAOI’s own leadership has been clear that pluggable transceivers, the current dominant approach to optical connectivity, hold the largest share of the photonics market today and aren’t expected to lose that position for years, even as co-packaged optics (CPO) grows alongside them. That matters because it means AAOI doesn’t need CPO to work out to keep growing, pluggables alone are described as a real, durable, multi-year business, with CPO as a second growth leg layered on top rather than something AAOI is betting its whole future on.

The China transceiver story just got a lot more interesting, and it’s very recent. On August 4, 2026, Reuters reported the FCC is drafting a rule to ban imports of new Chinese-made optical transceivers, citing national security concerns around data theft and malware risk. Chinese manufacturers, led by Zhongji Innolight at roughly 27% global market share, are estimated to account for roughly 60% of global optical transceiver production. If this rule goes through, US cloud companies would need to shift sourcing toward suppliers like Coherent, Lumentum, and by extension other US-based makers like AAOI.

I want to be precise about where this actually stands, because it’s easy to round a headline up into more certainty than it deserves. This is a proposed rule, not a finalized one. Officials hope to have it in place before the end of 2026, but the FCC could still modify or shelve it entirely. There’s also a real structural wrinkle worth knowing: one report specifically notes that banning Chinese transceivers could cut roughly 60% of current AI data center supply overnight, while the Western replacement suppliers themselves still depend on Chinese-sourced indium, the raw material underneath the lasers. So this isn’t a clean, simple “China gets banned, AAOI wins” story. It’s a real, developing policy that could meaningfully help AAOI and its peers, sitting alongside a genuine supply-chain complication that isn’t fully resolved. China has also signaled it could retaliate, and has a track record of doing so with rare-earth export restrictions in similar situations before.

My long-term valuation thesis. Before this China development, I thought AAOI could reasonably be a $70-80 billion company if the current order momentum and capacity expansion continue playing out. With the added tailwind of a potential China import ban pushing more US hyperscaler business toward American suppliers, if execution continues as I expect, I believe a $100 billion valuation is achievable. I’ll say directly why I keep coming back to a comparison: AAOI’s growth trajectory right now, small base, explosive order growth, a critical component inside a buildout everyone needs, reminds me of the kind of setup that turned SanDisk into a generational winner in storage. That’s my own comparison and my own opinion, not a guarantee, and AAOI going from roughly $10-11 billion today to $100 billion would mean becoming one of the largest pure-play optical component companies in the world. That’s a real, historically rare outcome, not a base case.

Why $SIVE is still in the portfolio, just weighted lighter for now

I bought my first $SIVE shares back in March, and even then, the thesis included a Jabil pluggables angle. At the time it was speculative. It’s since become real, which is part of why I’ve stayed in.

The original thesis, and how it’s evolved. My original idea was that Sivers becomes the “king of CPO” over time by acquiring its way up the value chain into a full, integrated optics company, something closer to what Lumentum already is. I still think that’s the long-term path. What’s changed is that the near-term picture has gotten stronger and more diversified than I originally expected. Sivers now has the confirmed Jabil pluggables relationship actually converting from speculative to real, and there have been hints of additional pluggables-side relationships scaling alongside it, which would bring in revenue meaningfully sooner than a pure CPO-only story would.

The CPO side is still the long-term prize, and it’s genuinely broad. Beyond Jabil, Sivers has real, disclosed relationships across GlobalFoundries’ SCALE platform (feeding into AMD’s future flagship chip), Ayar Labs (inside Nvidia’s NVLink Fusion ecosystem, with AWS committed to building its next chip on that same standard), and POET Technologies (with a real, SEC-documented deal structure behind POET’s own relationship with a company called Lumilens). Every one of these is a real, disclosed relationship. None of them has a named hyperscaler customer confirmed yet, that’s the honest state of things, covered in detail across the rest of this series. But the sheer number of real, separate paths into the same long-term opportunity is what makes me think of Sivers as having genuine depth, not a single bet. What keeps me invested isn’t one partnership, it’s the number of independent ways Sivers could win. Every few months another breadcrumb seems to connect back to the same company. Eventually, I think the market starts connecting those dots too.

The bonus layers. Sivers also has real, disclosed activity in satellite communications, automotive LiDAR moving toward production, and a funded defense and space partnership with BAE Systems and MIT Lincoln Laboratory. None of that is the core thesis, but it’s real, disclosed business sitting on top of everything else. I think of Sivers as genuinely layered, like an onion, pluggables revenue now, multiple CPO paths maturing over the next few years, and defense and space work generating real, if smaller, revenue in the background the entire time.

My long-term valuation thesis. If execution continues as I expect, I believe a $100 billion long-term valuation for Sivers is achievable, on the theory that it eventually acquires its way into becoming a full-package optics company. In the nearer term, I think $70 billion is a real, if aggressive, milestone if the M&A path plays out and even a couple of the CPO connections convert into real, named revenue. I want to be direct about how extreme that number actually is: Sivers is around $1.4 billion today, so $70 billion would mean a roughly 50x move. That is an exceptionally rare outcome for any public company, and I’m holding it as my personal long-term thesis, not something I expect to happen on a predictable timeline, or something I’d tell anyone else to plan their finances around.

Why the weighting is what it is

AAOI has orders landing today and a clearer, faster path to its next milestone. Sivers has more paths, but each one is earlier-stage and still needs a named customer to actually show up before the bigger numbers become real. That’s the entire reason for being heavier in AAOI right now and lighter in Sivers, not because I believe in Sivers less, but because the timeline to AAOI’s thesis playing out looks shorter to me than Sivers’ does.

The fun part: what this account is actually worth if the targets hit

Before I get into the risks, I want to show the actual math, because I think it’s more powerful as real numbers than as a vague “millions” statement.

At inception, this account was worth about $90,142. AAOI at $126 and SIVE at 5.82

Now here’s where it gets fun. If AAOI reaches my $100 billion target, based on its current share count, that implies a share price around $1,246. My 500 shares would be worth roughly $623,000. If Sivers reaches my $70 billion, based on its current share count, that implies a share price around $197. My 5,000 shares would be worth roughly $986,000.

Combined, if both targets hit, this account would be worth approximately $1.61 million.

I want to be completely clear about what this is. This is not a projection, and it’s not what I expect to happen on any specific timeline. It’s simple math applied to two price targets that are entirely my own opinion, run through today’s actual share counts, so you can see exactly what “over $1 million” really means in dollar terms rather than just as a headline. Whether it happens depends on everything below.

The honest risk section

Both of these are genuinely volatile, speculative positions, and I’d rather say that plainly than bury it. AAOI is unprofitable today and priced for growth that has to keep accelerating to justify where the stock sits. Sivers carries everything covered elsewhere in this series, real controversy around short-seller allegations, heavy dilution, and a stock still down more than 50% from its own all-time high. The China transceiver ban is proposed, not enacted, and could be delayed, watered down, or shelved entirely. A 50x move in Sivers or a roughly 9-10x move in AAOI, already itself worth $10 billion-plus, are both historically rare outcomes, not base cases.

This is also exactly why I’m not running this as part of my main, diversified portfolio. It’s isolated on its own account, tracked publicly, with a real timeline attached, by 2030 I think we’ll have a much clearer read either way, and if the thesis breaks along the way, I’ll make changes. I’m not committed to holding this no matter what happens. I hold both because I believe in the long-term thesis, not because I think this is a low-risk or predictable path to a specific dollar outcome, and I’d encourage anyone following along to think about position sizing relative to their own diversified portfolio the same way, not as a replacement for one.

Over the next years, every major update, every quarterly result, every thesis change, and every buy or sell will be documented publicly. Maybe this account never reaches $1 million. Maybe it blows past it. Either way, by 2030 we’ll know whether this thesis was right, and you’ll be able to follow every step along the way.

Breadcrumb 4 covered Jabil, a real but flat connection, where I could at least offer an informed technical guess about a likely customer based on form factor. This one doesn’t even have that. I want to walk through it honestly rather than force a prediction that isn’t actually supported by anything.

What O-Net and Enablence actually are

O-Net Technologies, based in Shenzhen, China, is one of the largest suppliers of optical communication devices, modules, and subsystems in the world, with a leading position in optical chips, silicon photonics, optical coatings, and optoelectronic packaging. Enablence Technologies, headquartered in Ottawa, Canada and listed on the TSX Venture Exchange, designs and manufactures planar lightwave circuit optical chips, with a wafer fab in Fremont, California, serving datacenter, telecom, AI, and vision applications. Neither is a household name, but both are real, established players in the optical components industry.

Where Sivers fits

On March 17, 2026, all three companies, Sivers, O-Net, and Enablence, announced a partnership to build an 8-channel external light source module, aimed at supporting co-packaged optics rollout in AI data centers and HPC systems. The roles are specific: O-Net serves as the ODM partner, integrating Sivers’ DFB laser arrays with Enablence’s NxN Star Coupler technology into one scalable module. The announcement was made at the OFC industry show in Los Angeles, with all three companies showing the technology at their respective booths.

Where the hyperscaler part comes in, or doesn’t

It doesn’t, not yet. The announcement describes the target market as AI datacenters and HPC systems broadly. No hyperscaler is named. No hyperscaler is even hinted at through an unnamed insider claim the way Lumilens’ customer was in Breadcrumb 3. This is a real, technical, three-way manufacturing partnership aimed at a real and growing market, full stop.

The market context that is real, even without a named customer

One thing worth knowing. According to market research firm IDTechEx, the overall co-packaged optics market is projected to exceed $20 billion by 2036, growing at a 37% annual rate from 2026 to 2036. In announcing this partnership, Sivers cited that same IDTechEx research and stated that ELS solutions specifically are expected to constitute roughly 10% of that market. I want to be precise about the sourcing there: the 10% figure comes from Sivers’ own press release citing IDTechEx, not something I personally verified inside IDTechEx’s paid report. It’s a third-party market estimate rather than company guidance, but it gives some real context for the category this partnership is targeting, even without a specific buyer’s name attached yet.

What this could mean for the market cap

The anchor is Tim Savageaux’s estimate at Northland Capital Markets, cited earlier in this series, that the external light source market this technology category serves is worth more than $1 billion a year. That’s a real, named, on-the-record analyst figure, not something I invented. Everything past that point is my own layered assumption: what share of that $1 billion the O-Net, Enablence, and Sivers combination could plausibly capture, given it’s competing against other real approaches in the same space, including the Ayar Labs and POET-Lumilens paths already covered elsewhere in this series, and then what share of that combined revenue actually flows to Sivers specifically as the laser supplier inside a three-way partnership.

I want to be as direct as I’ve been anywhere in this series about what this table actually is. There is no customer behind any of these numbers. There is no confirmed market share behind any of these numbers. This is a total addressable market, sourced to a real analyst, run through two layers of my own assumptions, market share and content share, stacked on top of each other. Treat this as the loosest, most illustrative table in this whole project, useful for picturing the size of the category, not for predicting what Sivers will actually earn from it.

What to watch for

Any disclosure from O-Net, Enablence, or Sivers naming a customer, even an unnamed-but-described one, would change this connection’s status immediately. Beyond that, watch for volume or revenue figures in future earnings calls from any of the three companies specifically tied to this ELS module, which would at least confirm the partnership is converting into real business rather than staying at the demonstration stage it’s at today.

Next…

Breadcrumb 6, the Fortune 100 wearable customer, the actual regulatory filing behind it, and the honest state of the Apple speculation.

This is my own opinion, not financial advice. I am not a financial advisor. Do your own research before making any decisions.

Every connection in this series so far has had some kind of trail leading somewhere, even when the ending was fuzzy. This one’s different, and I want to say that up front so you’re not waiting for a big reveal that isn’t coming. Jabil and Sivers have a real partnership. Nobody has said who’s actually buying the end product. I’ll give you my own best guess further down, but I’ll be upfront that it’s a weaker, shakier kind of guess than anything else in this series. To make up for the lack of a customer, I’m going to do something new: use another company’s real, disclosed deal to show you what this could actually be worth in dollars, if Jabil landed something similar.

What Jabil actually is

Jabil is a huge manufacturing and engineering company, one of the better-performing stocks in the S&P 500 this year. Think of it like this: Jabil doesn’t invent the laser, Sivers does that. Jabil builds the finished product around it, the same way a car company doesn’t necessarily invent its own tires, but knows how to build a great car using someone else’s. Jabil makes everything from electronics to medical devices to industrial gear, and it also has its own in-house engineering team for photonics work, so it’s not purely an assembly line.

Where Sivers fits

In April 2026, Sivers and Jabil announced a real deal. Jabil is building a next-generation data transceiver, basically a small device that moves huge amounts of data at very high speed, using Sivers’ lasers inside it. The big selling point is power efficiency: this thing uses 2.5 times less power than the alternatives. That matters a lot right now, because AI data centers aren’t just running out of space, they’re running out of electricity. An industry report cited in the announcement says these faster transceivers will make up 80% of the entire market by 2030, with 225 million of them shipped that year alone.

Does anyone know who’s actually buying it?

No. Not yet. The announcement just says “hyperscale AI data centers,” which is basically industry-speak for “big cloud companies,” without naming one. Nobody’s even dropped a hint the way Lumilens’ CEO did in Breadcrumb 3. This is a real, solid manufacturing deal aimed at a real, growing market. It just doesn’t have a name attached to it yet.

My one guess, and why you shouldn’t weigh it too heavily

Every other unnamed customer in this series traces back to somebody actually saying something, even informally, like a CEO posting on LinkedIn. This one doesn’t. Nobody connected to Jabil, Sivers, or any big tech company has said a word about who’s buying this. So what I’m about to say is 100% my own guesswork, not anything anyone told me.

Jabil’s transceiver uses a specific physical plug-in style called OSFP. Not every big tech company uses the same style. Meta has standardized on OSFP. Microsoft tends to use a different one. Google mostly builds its own thing entirely. So just based on which plug fits which company’s hardware, Meta is the most logical fit for Jabil’s product. That’s it. That’s my whole guess. It’s a reasonable hunch based on public information, not proof, and I don’t want you treating it the same as the more solid stuff elsewhere in this series.

Even without knowing the buyer, the market itself is big. Industry trackers think somewhere between 5 million and 20 million of these next-gen transceivers will ship in 2026 alone, and right now there isn’t enough supply to meet demand. None of that tells you Jabil specifically is winning big. It just tells you there’s real money on the table.

What a real deal like this is actually worth in dollars

Since I don’t have a Jabil customer to point to, let me show you something almost as useful: a real deal from a real competitor in the same space, so you can see what these numbers actually look like once they’re confirmed instead of guessed at.

A company called Applied Optoelectronics, ticker AAOI, landed its first-ever deal for this exact kind of next-gen transceiver on March 9, 2026. The order was worth more than $200 million, from one long-time big-tech customer. Around the same time, AAOI also picked up $124 million in separate orders for its current-generation product from another major customer, an initial $53 million order followed by a $71 million top-up. I’ll be honest that my sources don’t clearly say whether these are the same customer or two different ones, so I’m treating them as two separate real data points rather than guessing they’re connected. Either way, this tells you one relationship with one big tech company can be worth well over $100 million, sometimes over $200 million on its own.

Here’s a fun detail: the day AAOI announced that $200 million order, its stock jumped almost 16% and the company gained about $1.3 billion in value in a single trading session. That’s what happens when a real deal lands instead of just being rumored.

One more honest note before I run any numbers. That $200 million is one order, a single moment in time, not a guaranteed yearly paycheck. If a company can land an order that big once in a market this starved for supply, it’s reasonable to think more orders could follow. I’m not building that assumption into the math below because I have no responsible way to size it. Just know that if anything, the numbers coming up are more likely to undersell the real opportunity than oversell it.

One more thing worth knowing before the numbers: AAOI builds its own lasers in-house. It doesn’t buy them from anyone. That means every dollar of AAOI’s deal stays inside AAOI. Jabil works differently, it builds the finished transceiver but buys the laser from Sivers, so any deal Jabil wins has to get split between Jabil’s own cut and whatever Sivers earns as the laser supplier.

The scenario: what if Jabil did half of what AAOI does

Here’s the idea I want to walk through: what if Jabil eventually landed deals worth about half the size of what AAOI is doing in this same product category. I want to explain why that’s a fair number to use, not just a number I picked because it sounded good.

This isn’t about Jabil stealing AAOI’s actual customer. AAOI’s deal came from one specific company; Jabil would almost certainly be working with a different one. What makes “half of AAOI” reasonable is simply how starved this market is for supply right now. Demand for these transceivers is outrunning supply by two times or more, and a major player, Lumentum, is reportedly sold out all the way through 2028. When there isn’t enough of something to go around, a new, capable supplier doesn’t need to steal anyone’s lunch, there’s plenty of demand nobody’s currently able to fill.

I’m building this out in two stages: what it could look like right now, using AAOI’s actual $200 million order as the yardstick, and what it could look like once things mature, using an estimate of AAOI’s 2027 revenue. For how much of that money would actually flow to Sivers as the laser supplier, I’m using a range from real industry cost-breakdown data referenced earlier in this series: somewhere around 10% on the low end, up to 15% on the high end for a design that leans heavily on Sivers-style laser technology.

Right now, conservative case. Jabil lands deals worth half of AAOI’s real $200 million order, so $100 million. Sivers gets a modest 10% of that as its laser-supplier cut, $10 million. Using a 20x sales multiple, that’s $200 million added to Sivers’ value, pushing the company to around $1.6 billion.

Right now, aggressive case. Same $100 million Jabil deal, but Sivers gets a bigger 15% cut this time, $15 million. At a 30x multiple, that adds $450 million, pushing Sivers to around $1.85 billion.

2027, conservative case. Analysts think AAOI could be doing around $1.48 billion in revenue by 2027. If roughly half of that ends up being this specific next-gen transceiver, that’s about $740 million. Jabil doing half of that scale is $370 million, and Sivers’ conservative 10% cut of that is $37 million. At 20x, that adds $740 million, pushing Sivers to around $2.14 billion.

2027, aggressive case. Same $370 million Jabil scale, but Sivers gets its bigger 15% cut, $55.5 million. At 30x, that’s $1.67 billion added, pushing Sivers to around $3.07 billion.

Here’s the full picture side by side:

I want to say this plainly one more time, because a clean table like this can make something feel a lot more certain than it actually is. Jabil doesn’t have a named customer today. This whole model is built on Jabil someday winning business roughly on the scale of a real competitor’s real deal, which hasn’t happened yet and might never happen. Every number past AAOI’s own two disclosed figures, the $200 million order and the $1.48 billion 2027 estimate, is my own assumption. Use this as a way to picture the size of the opportunity, not as a prediction.

What I can say honestly: this deal is real, and it adds to the case that Sivers’ lasers are genuinely in demand across more than just one customer. But if you’re looking for the single connection most likely to send Sivers’ stock flying tomorrow, this isn’t it. The table above shows what it could become, not what it is today.

What to watch for

The thing that would change everything here is Jabil or Sivers actually naming a customer, even a vague, unofficial hint like the one Lumilens’ CEO gave in Breadcrumb 3. Until that happens, I’d treat this one as a real, solid piece of the story sitting quietly in the background, not something to watch day to day.

This connection has the most legally concrete dollar figure of any Breadcrumb so far, an actual SEC-filed warrant structure. It also comes with something the last two Breadcrumbs didn’t have: a real, recent reason for caution about the partner in the middle of the chain. I want to walk through both honestly.

What POET Technologies actually is

POET Technologies is a Toronto-based company that builds something called the Optical Interposer, a platform that integrates photonic and electronic devices onto a single chip using wafer-level manufacturing. POET doesn’t make lasers itself. It needs an external light source to complete its platform, which is where Sivers comes in.

Where Sivers fits

Sivers and POET have a real, disclosed partnership, first announced September 29, 2025, with additional details released in November 2025 and again in February 2026. It combines Sivers’ customized high-power DFB laser technology with POET’s Optical Interposer to build External Light Source, or ELS, modules for co-packaged optics. Prototypes were targeted for demonstration to customers in the first half of 2026, with production readiness targeted for the end of 2026.

Where Lumilens fits

In May 2026, POET and Lumilens announced a real supply deal. Lumilens placed an order with POET worth $50 million for optical engines, the light-processing hardware built on POET’s Interposer platform. That alone is a solid, real number to work with.

But there’s more to it, and this part is unusually well documented. The deal also includes something called a stock warrant, basically a promise of future POET shares given to Lumilens, filed directly with the SEC. How many of those shares Lumilens actually gets depends on how much it keeps buying from POET going forward, and the deal is structured so that could add up to more than $500 million in total purchases over five years if it all plays out. Because this detail comes from an actual SEC filing and not just a press release, it’s about as solid and well-documented as any single number in this whole series.

Lumilens itself is a real, funded company. It was founded in 2024, is based in Belmont, California, and has raised more than $130 million from well-known venture investors including Mayfield and Spark Capital. It builds some of its own chip technology too, not just relying on POET’s platform. The company is targeting early sample products for late 2026, with real production timed to match when big cloud companies start deploying their next generation of hardware in 2027.

The hyperscaler claim, and exactly where it comes from

Lumilens’ CEO, Ankur Singla, updated his LinkedIn to describe the company’s initial customer as one of the top three hyperscalers, specifically naming the field as AWS, Microsoft, or Google. I want to be precise about what this is and isn’t. It is not in the official POET-Lumilens press release. It is not in the SEC filing that discloses the warrant structure. It comes from the CEO’s own social media post, relayed and screenshotted by stock-focused accounts on X. That makes it more credible than an anonymous rumor, since it’s the actual founder describing his own company’s customer, but it is still not a corporate disclosure, and neither POET nor Lumilens has repeated the claim in any filing I could find.

POET’s recent partner trouble

Three weeks before the Lumilens deal was announced, POET had one of its most disruptive partner separations on record. In April, 2026, Marvell, which had acquired a company called Celestial AI, cancelled all outstanding purchase orders it had inherited from that acquisition. The stated reason was that POET had allegedly publicly disclosed purchase-order and shipping details in violation of confidentiality obligations. POET’s stock fell 46% in a single session, its sharpest drop on record. Separately, a short seller called Wolfpack Research disclosed a short position in POET that same month, calling it an “obvious stock promotion” and raising concerns about POET’s cash position and share dilution.

I want to be clear about why this matters here specifically, not just as generic bad news. If the allegation behind the Marvell cancellation is accurate, that POET disclosed confidential customer details it shouldn’t have, that’s a real, specific reason to weigh POET’s reliability as a partner before assuming the Lumilens relationship, or the Sivers relationship sitting behind it, will unfold smoothly. I’m not in a position to verify whether the confidentiality allegation is true or whether POET was in the right. I can tell you it happened, it’s recent, and it’s the kind of thing worth watching closely rather than ignoring because a bigger, shinier number showed up three weeks later.

What is not confirmed

The hyperscaler’s identity is not confirmed by anyone with the standing to confirm it, only narrowed to three names by the CEO’s own social post. There is no disclosed breakdown of how much of the $50 million order, or the potential $500 million over five years, specifically reflects Sivers’ laser content flowing through POET’s interposer versus other parts of the module. Sivers supplies POET, POET supplies Lumilens, Lumilens supplies an unnamed hyperscaler. That is three links in the chain, and the most concrete dollar figure we have is for the middle link, not for Sivers specifically.

What this could mean for the market cap, initial order and production ramp

This connection has more layers of separation from Sivers than either GlobalFoundries or Ayar Labs, and now it also carries a partner-reliability risk the other two don’t. Every number below assumes Sivers is capturing a reasonable share of value as the $500 million structure, if it materializes, flows backward through POET to Sivers. That share is not disclosed anywhere. It’s my own assumption.

Initial order base case. Sivers sees a modest share of the existing $50 million order flow through as its own laser revenue, around $5 million. At a 20x multiple, that is $100 million in added market cap, putting Sivers around $1.5 billion.

Initial order bull case. Sivers captures a larger share of that initial order, around $12 million. At a 30x multiple, that is $360 million in added market cap, putting Sivers around $1.76 billion.

Production ramp base case. The relationship matures toward the disclosed $500 million cumulative ceiling, but slowly and partially, and Sivers’ share of that settles around $30 million a year. At a 20x multiple, that is $600 million in added market cap, putting Sivers around $2 billion.

Production ramp bull case. The full $500 million cumulative structure actually plays out over its five-year window, averaging out to real annual revenue for POET, and Sivers captures a meaningful share of that as the laser supplier underneath it, plausibly $80 to $100 million a year. At a 30x multiple, that is $2.4 to $3 billion in added market cap, putting Sivers around $3.8 to $4.4 billion from this connection alone.

Here is the full picture side by side:

I want to be extra clear. These numbers rest on more assumptions stacked together than either of the last two Breadcrumbs, not just revenue and multiple, but also what share of POET’s revenue actually becomes Sivers’ revenue, which nobody has disclosed, and now also whether POET itself executes cleanly on a major partnership after a public, contentious partner separation just weeks earlier. Treat this as the most speculative table in the series so far, and I mean that more than I meant it for any connection before it.

What to watch for

The single biggest thing that would change this thesis is Lumilens, POET, or the hyperscaler itself confirming the customer’s identity in an actual filing or press release, not a LinkedIn post. Beyond that, watch for how POET’s relationship with Lumilens is discussed on POET’s own earnings calls, whether any further confidentiality or contract disputes surface, and whether POET discloses any Sivers-specific revenue tied to this order as it progresses.

Next…

Breadcrumb 4 moves to Jabil, a quieter connection, real partnership aimed at AI data centers that hasn't yet produced a named customer of any kind, confirmed or claimed.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

This connection has developed faster than any other in this series, and it involves the two biggest names of all six: Nvidia and Amazon.

Let's recap today's market cap check before we start. $SIVE is trading at about 47 SEK per share today, for a market cap of roughly 14 billion SEK, or about $1.4 billion USD, still down around 57% from its all-time high of 110 SEK per share, or roughly $3.36 billion in market cap, reached back on June 3, 2026.

This piece is about the chain that runs from Sivers, through a company called Ayar Labs, into Nvidia’s own connectivity standard, and toward AWS. Three of those four links are confirmed. One is not.

What Ayar Labs actually is

Ayar Labs builds the optical engine, the piece of hardware that turns electrical signals into light and back again. If Sivers makes the laser, Ayar Labs makes the box the laser sits inside of. Ayar Labs closed a $500 million funding round in March 2026 at a $3.75 billion valuation, with Nvidia and AMD both listed as strategic investors. Their CEO has said this round is likely their last before a possible IPO. This is a well-funded company sitting right at the center of the optical I/O conversation, not a garage startup.

Where Sivers fits

Sivers has supplied Ayar Labs with laser arrays since 2022, first shown publicly at an industry conference that year. In July 2023, Sivers confirmed a $1 million order for next-generation laser arrays, including agreed pricing terms for future volume. In December 2024, the relationship expanded again around manufacturing readiness for high-volume deployment. Ayar Labs’ own public roadmap points to product qualification by the second half of 2027, with real volume ramp expected around 2028.

Where Nvidia fits

In June 2026, Ayar Labs officially joined Nvidia’s NVLink Fusion ecosystem. NVLink Fusion is Nvidia’s open standard for how chips and systems connect to each other, and Nvidia lets outside companies build compatible parts for it instead of keeping everything proprietary. Being part of that ecosystem means any hyperscaler building Nvidia-compatible AI infrastructure can choose to use Ayar Labs’ optics as part of the design.

Worth knowing here too. Nvidia has announced over $6 billion in investments and strategic supply agreements across the photonics ecosystem since March 2026, including $2 billion each into Coherent, Lumentum, and Marvell, up to $3.2 billion into Corning, and participation in Ayar Labs’ $500 million round specifically. Ayar Labs is one of the names Nvidia itself chose to back.

Where AWS fits

AWS has announced Trainium4, along with its Graviton CPUs and Nitro system, as part of its multigenerational NVLink Fusion partnership with Nvidia, unveiled at AWS’s re:Invent conference in December 2025. This comes straight from Nvidia’s own developer blog. In concrete terms, NVLink Fusion lets a rack connect up to 72 custom chips together as if they were one large computer, moving data between them at up to 260 terabytes per second, and Nvidia has said the approach can deliver up to three times the performance for AI inference compared to the previous generation. That is a real, named hyperscaler, publicly committing to a real, named standard, for real technical reasons, that Sivers’ biggest optics customer already belongs to.

So the chain looks like this. Sivers builds the laser. Ayar Labs builds it into the SuperNova light source that feeds its optical engine. Ayar Labs joins Nvidia’s NVLink Fusion standard. AWS builds its next chip on that same standard. The first three links, Sivers to Ayar Labs to Nvidia’s ecosystem, are confirmed. AWS’s commitment to the standard itself is confirmed too. The one piece still missing is AWS actually naming Ayar Labs, specifically, as its optics vendor inside that standard.

Where this chain currently stops

NVLink Fusion has more than one optics partner. Alongside Ayar Labs, both Lightmatter and GlobalFoundries’ own SCALE platform are also part of that ecosystem. AWS has confirmed it is building Trainium4 on the NVLink Fusion standard. AWS has not said which optics technology it will actually use inside that implementation. Ayar Labs is one of at least three real options sitting in that ecosystem, not the only one. AWS also has not given a launch date for Trainium4, unlike AMD’s MI500, which has a stated 2027 target.

Something worth understanding about Silicon Photonics here

Ayar Labs’ technology is built on Silicon Photonics, chips made of silicon that handle light instead of electricity. A Goldman Sachs report on optical networking projects Silicon Photonics adoption in the broader datacom market jumping from 6% in early 2024 to 45% by late 2028. That might sound like a threat to a laser company. It probably is not. Silicon itself cannot efficiently create light on its own, so every Silicon Photonics chip still needs an external laser fed into it from outside, which is exactly the role Sivers plays for Ayar Labs. As Silicon Photonics adoption grows, the need for external laser sources feeding those chips should grow right alongside it, not shrink.

The same report also broke down the actual cost structure of a real, comparable CPO switch. Optical engines, the category Ayar Labs itself operates in, made up 42.7% of the entire component cost of that switch, by far the largest single piece. That gives a real, third-party sense of how valuable the position Ayar Labs is fighting for actually is.

What this could mean for the market cap, initial order and production ramp

Everything below depends on two things happening, not one. Trainium4 has to actually ship, and Ayar Labs specifically has to be the optics vendor AWS picks over Lightmatter and GlobalFoundries. Neither is confirmed yet. This connection carries real vendor-selection risk that Breadcrumb 1 does not have, since Sivers’ laser is the default component in GlobalFoundries’ blueprint that AMD is building on, though AMD could in theory swap in a different supplier, whereas with AWS, Sivers still has to win the optics slot outright against named competitors.

Here’s the flip side. NVLink Fusion is not just about AWS. Any other hyperscaler or chip designer building on that same standard could also choose Ayar Labs, the same way AWS is considering it. None of the numbers below account for that.

Nobody, not Sivers, not Ayar Labs, not Nvidia, not AWS, has said how many dollars of laser go into this, or what volume is expected. Every number below is my own estimate, not something any company has confirmed.

Initial order base case. Ayar Labs is selected as one of possibly several approved optics vendors for early Trainium4 deployments, and Sivers sees modest revenue from it, around $8 million annually. At a 20x sales multiple, that is roughly $160 million in added market cap. On top of today’s $1.3 billion, that puts Sivers around $1.5 billion.

Initial order bull case. Ayar Labs establishes an early, meaningful foothold in AWS’s first Trainium4 deployments, and Sivers sees around $15 million annually tied to it. At a 30x sales multiple, that is roughly $450 million in added market cap, putting Sivers around $1.8 billion.

Production ramp base case. Trainium4 reaches steady, real production volume, and Ayar Labs holds a real but not dominant share of AWS’s optics business. Sivers’ revenue tied to this settles around $40 million a year. At the same 20x multiple, that is roughly $800 million in added market cap, putting Sivers around $2.1 billion.

Production ramp bull case, the rack-scale story. This one needs its own explanation, because it’s a genuinely different scale of opportunity than switch-level CPO alone. At Computex 2026, Ayar Labs and Wiwynn showed a rack-scale reference design where SuperNova light source modules, the ones housing Sivers’ lasers, are packed in at a density well beyond the earlier switch-centric picture. Publicly shown system diagrams pointed to up to 16 SuperNova modules per compute tray. One independent industry analyst, extrapolating that across a full rack of roughly 32 trays, calculated a theoretical maximum of around 512 SuperNova assemblies per rack, a figure worth being precise about, since it’s that analyst’s own calculation from public diagrams, not a number Ayar or Wiwynn has stated directly. If rack-scale architectures like this move into real hyperscale production, the number of light sources needed per rack could end up well above what most switch-focused CPO models, including Goldman’s, currently assume.

Here’s roughly how that could play out for Sivers. Using an illustrative assumption of roughly $60 to $65 of Sivers content per SuperNova assembly, my own estimate rather than a disclosed figure, large-scale deployment across many racks could plausibly turn into laser revenue somewhere in the $125 million to $205 million a year range once things reach real hyperscale volume. Using the same 30x multiple as the rest of this series, that is $3.75 to $6.15 billion in added market cap, putting Sivers somewhere around $5 to $7.5 billion from this connection alone. Worth being direct about how big a scenario this actually is too, that’s Ayar Labs alone growing to somewhere between 4 and 6 times Sivers’ current total revenue. It is a real, technically grounded scenario if rack-scale density holds up at hyperscale volume, not a base case.

Here is the full picture side by side:

None of this is guidance from Sivers, Ayar Labs, Nvidia, or AWS. These are my own assumptions. This connection carries a real competitive risk that the GlobalFoundries and AMD relationship does not, since Sivers still has to win the optics slot rather than being in the blueprint by default, but the rack-scale density story is also a genuinely different scale of upside than anything in Breadcrumb 1, which is why the bull case here ends up in a similar range despite that added risk.

Next…

Breadcrumb 3 moves to a smaller company most people have never heard of, and a $50 million order that comes with a very specific, very real claim about landing a top-3 hyperscaler as a first customer.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

This connection is the one with the clearest paper trail of all six, and it starts with a company most people confuse for something it is not.

Let’s recap today’s market cap check before we start. $SIVE is sitting at about 12.6 billion SEK today, roughly $1.3 billion USD, still down around 61% from its all-time high of 110 SEK per share, or roughly $3.36 billion in market cap, reached back on June 3, 2026.

This piece is about one specific chain that runs from Sivers, through GlobalFoundries, into an AMD chip, and toward the hyperscalers who would actually buy it. Three of those four links are confirmed. One is not. Here is each step, in order.

What GlobalFoundries actually is

GlobalFoundries is a foundry. That means it is a contract manufacturer. It does not design chips and it does not sell finished products to anyone. Think of it like a factory that other companies hire to build the thing they designed. GlobalFoundries builds chips for other companies, the same way a company like Foxconn builds iPhones for Apple without ever putting its own name on the box.

In June 2026, GlobalFoundries announced something called SCALE. This is not a product you can buy off a shelf. It is a reference design and a manufacturing process, basically a blueprint plus the factory capability to build it, specifically for something called co-packaged optics, which is the industry term for putting the optical connection right next to the chip instead of running a separate cable to it.

Where Sivers fits

Sivers’ laser arrays are one of the components built into that SCALE blueprint. Sivers' lasers are part of GlobalFoundries' SCALE ecosystem, which means they are available to any chip designer adopting the platform. Sivers does not need to go sign a deal with every single chip company in the world. It just needs its laser to stay part of the blueprint that chip companies are choosing to build on.

Where AMD fits

AMD designs chips. It does not manufacture them itself for this purpose, it hires foundries like GlobalFoundries to do that. AMD has confirmed it is using GlobalFoundries’ SCALE platform for the optical interconnect portion of its Instinct MI500 chip, a future flagship targeted for 2027, two generations ahead of the MI300X and MI400/MI450 chips currently shipping.

So the actual chain looks like this. Sivers builds the laser. GlobalFoundries builds it into their SCALE blueprint. AMD designs a chip on that blueprint. AMD then sells that chip to hyperscalers, the cloud giants who actually run AI data centers. The first three links, Sivers to GlobalFoundries to AMD, are confirmed. That chip is confirmed. That timeline is confirmed. AMD is a real, named company, not a guess. The fourth link, AMD to a specific hyperscaler, is the one piece still missing.

Where the hyperscalers fit, and where this chain currently stops

AMD is not a hyperscaler. AMD designs chips and sells them to the cloud giants, companies like Microsoft, Amazon, Google, and Meta, who actually run the data centers. Nobody has confirmed which of AMD’s hyperscale customers will be buying MI500 chips specifically. What we do know is that AMD’s existing Instinct chips are already running at Microsoft Azure, Meta, Oracle, and inside OpenAI’s infrastructure. AMD has a strong track record of its existing customers adopting each new Instinct generation as it launches. If that pattern holds for MI500, that is how this chain, Sivers to GlobalFoundries to AMD to hyperscaler, finally closes. It has not happened yet.

Separately, GlobalFoundries’ own CEO has said the company is designed into three of the top four pluggable optical transceiver companies in the world. He did not name them. That is a second, real path from Sivers to a hyperscaler, running in parallel to the AMD chain, and it also stops short of naming names.

What this could mean for the market cap, initial order and production ramp

Now, before any numbers, something important. Everything below is based on one single chip, AMD’s MI500. Not AMD as a whole company. Not GlobalFoundries as a whole company. One product that has not even shipped yet. If MI500 gets delayed past 2027, this whole timeline shifts with it. If it gets cancelled, or redesigned in a way that drops this light-based approach, this specific connection could end up worth very little to Sivers, even if the relationship with GlobalFoundries itself stays solid.

This section only counts what we can confirm today, the Sivers to GlobalFoundries to AMD chain, tied to MI500 alone. That likely undercounts the real opportunity. GlobalFoundries has separately said it is built into three of the top four transceiver companies, without naming them, so other products could already be using Sivers through GlobalFoundries that we just do not know about yet. And AMD will keep making chips after MI500, likely on the same SCALE platform, the same way any other chip company that adopts SCALE would inherit Sivers too. None of that is in the numbers below.

One more honest note before the numbers. Nobody, not Sivers, not GlobalFoundries, not AMD, has said how many dollars of laser go into one MI500 chip, or how many chips AMD expects to sell. So every number below is just my own estimate, not something any company has confirmed. Think of it as a way to picture the size of the opportunity, not a prediction.

Initial order base case. This means MI500 ships on schedule in 2027 and starts selling slowly at first, the way most brand-new chip designs do. Sivers’ laser business picks up a real but modest amount of extra revenue from it, around $10 million a year. Here is where it gets interesting. Sivers already trades at a high multiple today, somewhere around 30 to 50 times its current yearly sales, even without this revenue confirmed. If the market applies a 20x multiple to that new $10 million once it becomes real and repeating, that adds roughly $200 million to the company’s total value. On top of today’s $1.3 billion, that puts Sivers around $1.5 billion just from this one connection getting started. Not the whole story, but a real, visible step.

Initial order bull case. This is the stronger version of the same idea. In MI500’s first year or so of shipping, before volume has time to fully build up, an unusually strong start could plausibly bring in around $20 million a year for Sivers. Apply a 30x multiple, which is not unusual for a supplier proven to be inside a big, named AI chip while the story is still fresh, and that adds roughly $600 million in value. Stack that on today’s price and you get something around $1.9 billion in this early, best-case scenario, before any of the other five connections in this series add anything at all.

The bigger number might come sooner than I first thought

Here is an honest timing question. This is a brand new partnership, and deals like this do not go from announcement to full-scale sales overnight. But I want to correct myself a bit here, because when I actually looked at industry data instead of assuming a generic multi-year wait, the timeline is tighter than I first thought.

Multiple industry reports on this kind of optical technology point to the same window: real, large-scale use in AI chips is expected to start in 2027 and hit high volume by 2028, once supply chains and manufacturing catch up. That lines up almost exactly with when MI500 itself is supposed to ship, with real volume following about a year later. So the bigger, more mature scenario below might only be twelve to eighteen months behind the early scenario above, not the multi-year gap I originally pictured.

Production ramp base case. This is what it could look like once things are running smoothly, not just starting out. MI500 becomes a steady, ongoing part of AMD’s lineup, not a runaway hit, just a solid seller. Sivers’ laser revenue tied to it settles around $50 million a year. Using the same 20x multiple as above, that adds roughly $1 billion in value. On top of today’s $1.3 billion, and assuming the rest of the business holds steady, that puts Sivers around $2.3 billion.

Production ramp bull case. This is the strong version of that same mature outcome. MI500 becomes one of AMD’s flagship products, selling in the millions of units a year the way AMD’s best chips already do, and Sivers becomes the main, go-to laser supplier for it. Revenue tied to this could plausibly reach $150 to $200 million a year at that point. Using the same 30x multiple, that is somewhere between $4.5 and $6 billion in added value. Stack that on today’s base and Sivers could be worth somewhere between $5.8 and $7.3 billion from this one connection alone, before counting any of the other five connections in this series. That is real, meaningful progress toward the ten billion dollar case from the introduction, on a timeline that might be closer than a typical multi-year small cap story.

Here is the full picture:

Again, none of this is guidance from any company involved. These are my own assumptions about how much revenue could show up and what multiple the market might pay for it, meant to illustrate what a fully-matured outcome could look like, not to predict an exact number. The real point is that the payoff here plays out over a couple of years, not next quarter, and that is exactly why I think the market has not priced it in yet.

2027 is around the corner. It lines up with when the broader industry itself expects this technology to hit real volume, not just with when MI500 happens to ship. If AMD gives any update on who is buying MI500 chips, or if GlobalFoundries announces new customers on its SCALE platform, that is when I would expect this specific connection to move the stock the most, even before the actual revenue shows up.

Next…

Breadcrumb 2 moves to the connection I think has developed the fastest in just the last few months: Ayar Labs and Nvidia’s NVLink Fusion ecosystem.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.

$SIVE is sitting at about 12.6 billion SEK right now. That is roughly $1.3 billion USD. That is also down about 61% from its all-time high of 110 SEK per share, or roughly $3.36 billion in market cap, reached on June 3, 2026. Here is the thing that gets me. That price, even after that drop, is for a company where none of its biggest partnerships have actually turned into real volume revenue yet. Zero of them. Every relationship I am about to walk you through is confirmed to exist and confirmed to be moving toward production, but almost none of it has hit real revenue on the AI and hyperscaler side yet.

If even one of these connections partially executes, meaning it goes from confirmed relationship to actually shipping in real volume, I think this valuation looks silly in hindsight. Not all six connections. Not even most of one. Just one connection starting to convert into real numbers the market can point to.

This is my own opinion, not a guarantee, but I think a $10 billion market cap is realistic in the short term if any single connection here starts turning into real revenue. That is roughly a 7.7x from where we sit today. I am not saying it happens next quarter. I am not saying it is guaranteed. But that gap, between one connection starting to execute and where the stock trades today, is why I have not been this excited about a small cap in a long time.

Here is the obvious question that deserves a direct answer. If this is real, why isn’t the stock already at $10 billion. The market generally discounts relationships like these until they actually turn into purchase orders and recurring revenue, not when they are merely disclosed. My thesis is not that these six relationships exist. They already do, and that part is not really in dispute. My thesis is that the market is underpricing what happens once even one of them converts into sustained, recurring production revenue instead of a disclosed partnership sitting on paper.

That is the whole thesis of this series in one paragraph. Now let me show you why.

Six connections, one small Swedish laser company

I could honestly write forever about these partnerships. This one small Swedish laser company has six real, confirmed connections sitting one or two steps away from some of the biggest hyperscalers and chip companies in the world.

That is what this series is about. I have spent a lot of time going through every filing and press release I could find, tracing where each partner’s technology actually ends up, and figuring out what is confirmed versus what is just community guessing. This series is going to walk through each of the six connections, one at a time, and every part is going to come back to that same question. Is this the connection that closes the gap between $1.3 billion and $10 billion.

Who is Sivers?

Sivers Semiconductors is a Swedish photonics and wireless company, ticker SIVE on Nasdaq Stockholm, built around something genuinely useful: high precision laser chips. Photonics is the segment that matters most for this series. Wireless (the radio chips used in 5G, 6G, and satellite communication) is real too, but it is mostly a separate story. The company does around $33 million a year in revenue today, against a disclosed pipeline of potential future business that grew 77% in the first five months of 2026 alone, hitting $799 million. That is a small company with a pipeline more than twenty times its current revenue. That gap is the whole reason this stock has people watching it closely, and it is exactly the gap behind the valuation case above.

The problem Sivers is actually solving

To understand why any of these six connections matter, it helps to understand the physical problem sitting underneath most of them.

For years, the chips inside an AI data center, the GPUs and custom accelerators doing the actual training and thinking, have talked to each other over copper wires. Same basic idea as a phone cable. That works fine at small scale. But modern AI clusters connect thousands of chips together, and copper hits hard physical limits. Past a certain distance and speed, it needs more and more power just to push a signal through, it generates heat, and it simply runs out of bandwidth. As AI models get bigger and clusters grow from thousands to hundreds of thousands of chips, copper is turning into the bottleneck that limits how fast the whole system can run. Not the chips. The wires.

The fix the whole industry is racing toward is swapping those copper connections for light. Sending data as pulses of light instead of electrical signals means way more bandwidth, over longer distances, using less power. This shows up under a few different names depending on exactly where the optics sit relative to the chip (co-packaged optics, near-packaged optics, external light sources), but the core idea is always the same: get light doing the job copper used to do.

Light needs two things copper does not: a laser to create it, and an engine to turn electrical signals into light and back again. Sivers makes the first part, high precision laser chips. It turns out a surprising number of the companies building the second part, the optical engines, all rely on Sivers as their light source. That is the common connection running through five of the six connections in this series. Different partners, different chip designs, different customers, all needing the same basic component to make optical connections actually work.

The six connections, quick preview

GlobalFoundries to AMD. Sivers’ lasers are built into GlobalFoundries’ co-packaged optics platform, which AMD is using for its upcoming Instinct MI500 chip, targeted for 2027. A real foundry relationship feeding a real, named chip, with the eventual hyperscaler customer still unnamed.

Ayar Labs to Nvidia NVLink Fusion to AWS. Sivers’ laser powers Ayar Labs’ optical engine. Ayar Labs sits inside Nvidia’s NVLink Fusion ecosystem, the same standard AWS just confirmed it is building its next chip, Trainium4, on. Which optics vendor AWS actually uses is still an open question.

POET Technologies to Lumilens. A confirmed Sivers and POET laser partnership, sitting one step behind POET’s real $50 million order from a startup whose stated first customer is a “top-3 hyperscaler,” named by people close to that startup, not by POET.

Jabil. A 2026 partnership on next-gen pluggable transceivers for AI data centers.

O-Net and Enablence. A three-way partnership building external light source modules for AI data center and high performance computing applications.

The Fortune 100 wearable customer. A real relationship, real dollar figures, unnamed identity. The community has been pointing at Apple, with a competitor’s very public collapse (a company called Rockley Photonics) sitting in the background of that theory.

There is also real activity outside these six worth knowing about: the ALL.SPACE satellite program, the Wireless segment more broadly (including relationships like Nokia on 5G mmWave beamforming and Tachyon Networks on fixed wireless access), the LiDAR business moving toward series production, and a funded defense partnership with BAE Systems and MIT Lincoln Laboratory. All of this is noteworthy and part of why I am bullish on this company overall. But I am not going to be covering it in this series. This series is specifically about the six connections that lead toward hyperscalers and AI infrastructure spend, so that is where I am keeping the focus.

One more thing. I am not the first person talking about this stock and I am not going to pretend I am. There is already a community around it. What I am trying to do here is put every real connection in one place, explain it like you have never heard of any of this before, and give you my honest take on what it might be worth once you have seen the whole picture.

What’s coming

The first breadcrumb is GlobalFoundries and AMD, the connection with the clearest paper trail of all six.

This is my own opinion, not financial advice. I am not a financial advisor and this is not a price target you should trade on. Do your own research before making any decisions.