Reading Notes: This summary is compiled in accordance with the internationally recognized Chatham House Rule—it only presents viewpoints without pointing to any specific speaker or institution; third-party company names mentioned during the discussion are retained as they are (the rule protects "who said it," not "who was talked about"); all figures in the text are impromptu expressions by on-site experts, not independently verified, and should be used as "frontline judgments" rather than "confirmed facts"; the entire text does not constitute any investment advice.
01. Why Hold This Meeting Behind Closed Doors
On July 10, in Shanghai, on a sweltering summer afternoon. Outside the venue was the oppressive summer heat, but the heat inside was even higher—over a hundred people registered publicly, but only a little over forty finally sat in this meeting room. There was only one screening criterion: Are you on the frontline of the supply chain, and do you have first-hand information?
This was not a forum. No PPTs, no speeches, only voting cards, roll calls, and follow-up questions. Over forty frontline figures in the silicon photonics supply chain faced off for four and a half hours around six of the most divergent issues in the industry.
Why hold it behind closed doors? At the opening, the organizer shared an industry anecdote: Over fifty years ago, the global semiconductor industry suffered greatly from non-unified standards—interfaces among numerous equipment and material companies were fragmented, leading to the establishment of the industry organization SEMI. Over the next fifty years, through numerous similar closed-door seminars, SEMI formulated various basic standards for today's semiconductor industry: interconnection standards between equipment, and even why wafers are 4-inch, 6-inch, 8-inch, and 12-inch, rather than having 9-inch and 9.5-inch like pizzas—these were all reached by frontline experts at closed-door meetings and then published as industry standards by SEMI, profoundly influencing the development of the semiconductor industry over the past fifty to sixty years. The most important consensuses were mostly born in quiet rooms, not on lively podiums.
Silicon photonics is now reaching a similar crossroads: scarce capacity, route divergences, and a mix of true and false information. With AI, there is more noise than ever before; while the traffic is in the lively places, the value lies in the quiet places.
02. The Rules of This Meeting: Chatham House Rule
The entire session strictly applied the Chatham House Rule, universally used for international closed-door meetings, and was signed by each participant. This rule was born in 1927 at the Royal Institute of International Affairs in the UK, and its core is just one sentence: Participants are free to use the information and views received from the meeting, but must not reveal the identity and affiliation of the speaker or any other participant.
The purpose of the rule is not confidentiality, but to empower those with genuine information to speak the truth—views can be taken away, but names must be left behind.
The process design itself is also worth mentioning: each topic is first "blind-voted" by the whole room, followed by free debate, and then re-voted after the debate—being persuaded to change sides is not embarrassing; on the contrary, it is the most valuable data. Speaking is limited to 90 seconds, and the moderator holds the written positions submitted by each participant before the meeting. When the room goes cold, the moderator calls on them: "You didn't vote for this before."
More importantly, there is a mechanism commitment: all judgments will be reviewed on a quarterly basis, and answers will be checked item by item after three months—those who are right are prophets, and those who are wrong will review together. Next session, October.
03. Who is Sitting at This Table
In accordance with the rule, we will not publish the attendee list, but we can provide a complete anonymous profile—this table is almost fully occupied by representatives from every segment of the silicon photonics supply chain, and they are all "people with first-hand information":
One detail illustrates the caliber of this list: during the self-introduction session, at least six guests independently pointed out the same pain point from their respective perspectives—"The CPO route, process, and collaboration methods have not yet converged." This is exactly the background of these six debates.
04. Six Questions and Six Answers: Real Tendencies On-Site
Question 1 | Will Silicon Photonics Capacity Be in Oversupply Within Three Years?
On-site conclusion: Refusal to give a simple answer—"oversupply" and "shortage" are simultaneously true on the same chain.
The most unexpected part was not the answer, but the vote: about 80% of the participants refused to express their stance. The collective prudence of frontline figures is itself important data.
The arguments for oversupply come from history: photovoltaics, lithium batteries, mature processes, and silicon carbide serve as recent lessons—"standard products that China can manufacture at scale will inevitably be in oversupply"; under the arms race, major US cloud providers spend one to two hundred billion dollars in capital expenditure annually, and "each will build more than they need."
The non-oversupply camp presented capacity projections: according to Tower's expansion plan, the visible capacity by the end of 2027 corresponds to about 80 million to 100 million 800G equivalent optical modules—while the demand this year is about 65 million, and next year 150 million to 200 million. "Based on the currently visible capacity, there is a 99% probability it will not be in oversupply."
The discussion ultimately converged on a layered judgment: the simple pluggable segment "will be in oversupply as long as China can make it," and some are already in oversupply; upstream optical chips, EML, and silicon photonics wafers are instead in short supply—the on-site consensus is a gap of about 30% to 40%, lasting at least until next year, and "the bottleneck changes every two quarters"; CPO capacity is far from sufficient.
There is also a perspective on the industry landscape: the biggest threshold in this industry has never been capacity, but customer resources and positioning. Google's optical module suppliers have historically never exceeded 5; top manufacturers form structural cost advantages by binding chip resources—on-site consensus: InnoLight binds with Marvell, Eoptolink binds with Broadcom, TFC Optical binds with NVIDIA; for the same DSP, the purchase price for top manufacturers like Eoptolink is about $50, while a second-tier manufacturer with a tens-of-billions market cap has to pay over $70—"Even if I get the order, I still can't make money." Top players continuously making money while second- and third-tier players continuously losing money is the norm in this industry.
A supply chain entrepreneur who prepaid a deposit for 2028 capacity added the most realistic footnote to this question: "I bet right last year, but this year I'm getting more and more nervous—every penny paid now is to secure a spot for 2028; once the deposit is paid, the full payment must be made."
The greatest common denominator is one sentence: the only posture to resist oversupply is to stand at the technological frontier—"What we fear is not oversupply, but that it becomes obsolete just as we finish making it."
Question 2 | Is NPO a Transition or the Correct Answer for China's Supply Chain?
On-site conclusion: Highly converged—NPO is at least the correct choice for China at this stage, and "transition" and "answer" are not contradictory.
The reasons are closely linked: China has the largest optical module capacity in the world, and CPO will swallow the existing value of module manufacturers; China lacks the TSMC (Taiwan Semiconductor Manufacturing Company)-level large-scale advanced packaging and high-end electronic chip ecosystem, "and cannot go the CPO route when severely lacking electronic chips." The US is exactly the opposite—without an optical module supply chain like China's, it leans towards CPO. The route divergence is structural: both sides are choosing the route that conforms to their own industrial endowments.
The evidence on the demand side is equally solid: overseas top cloud providers like Amazon and Meta have signed long-term agreements for NPO, with volume ramping up next year; next year's mainstream NPO light source will be about 120mW; NVIDIA is testing 3.2T NPO, and Amazon is testing 6.4T NPO (on-site consensus); Tower's new factory coming into production in 2028 is "basically built entirely for NPO"; the NPO promotion organization led by Huawei was established just the day before the meeting.
The sharpest viewpoint in the room pulled the problem into another coordinate system: whether CPO can be adopted is not decided by the optical communication industry, but by AI chips. "Whoever the AI chip chooses is the only one qualified to do CPO with them; otherwise, you don't even have the qualification to sit at the table." "Fast abroad doesn't mean fast at home—it's not that optical communication experts are inadequate, but that the domestic AI chip ecosystem is still too weak." For most domestic GPU companies, "before talking about CPO, first talk about whether you can survive."
There is also a widely underestimated reminder: the O&M convenience of NPO should not be overestimated either. In reality, if several NPOs in a row of cabinets have problems, the whole machine still needs to be pulled out for repair; the signal integrity, impedance, and stress of the socket are hard problems—"Everyone thinks you just plug it in and it works, but it's actually not."
Historical lessons are repeatedly cited: Intel, the silicon photonics leader with the most advanced technology ten years ago, almost died later. "Too early never works."
Question 3 | Where Are the Shortcomings of Domestic Silicon Photonics, in Optics or Electronics?
On-site conclusion: Almost unanimous—it's in electronics. But "materials" emerged as a third answer.
The evidence chain for electronics is densely shocking: globally, only Marvell, Broadcom, and NVIDIA can mass-produce and deliver high-speed DSPs; for 1.6T modules, basically only Marvell's TIA can pass the verification of top customers, "the value is not high but it's a choke point"; the high-end segment of SerDes is nearly blank, "112G will be stuck for a while, 224G power consumption can't come down, and going further up will definitely be stuck"; even high-end capacitors have faced shortages.
"Optics," on the contrary, has already broken through or is running neck and neck. Why? One participant gave the most thorough explanation in the room: "Moore's Law fails in the manufacturing of optics—this is precisely the most friendly place for domestic players." Optical manufacturing does not rely on advanced processes; the process stays in place waiting for latecomers to catch up; five to eight years ago, China already occupied seven of the top ten global optical module companies; the silicon photonics chips used by InnoLight's 800G/1.6T have all been self-developed (on-site consensus).
Why is electronics difficult? "Revising a complex electronic chip costs about $100 million each time; without manufacturing capability, you cannot hone design capability—chips are made through iteration." How many times can a startup revise?
But there is more than one way to break the deadlock. First, detour: use an indium phosphide (InP) substrate to strip away the parts that must use advanced processes, "the industry is far from reaching a bottleneck, it's just that no one has tried it yet." Second, collaboration: optoelectronic packaging collaborative design—"after adding packaging factors, optical modules are no longer standard products," and collaborative design among optics, electronics, and packaging is an opportunity to overtake on the curve. Third, change materials: the public data of thin-film lithium niobate (TFLN) performance is better than silicon photonics modulators (referring to the public data of Harvard-affiliated HyperLight products), yet the industrialization threshold is extremely low—the account given on-site is that the ultimate material cost for an 8-inch wafer is about 2,000 RMB, and the entire line investment is about 100 million RMB; domestic turnkey solutions already exist for the two key equipment of ion implantation and hybrid bonding, "the threshold is so low that even individual entrepreneurs can do it." The acoustic filter industry has already proved this once with the same script: "A change in materials will bring about a change in the industry." New strained germanium substrates (with hole mobility reaching the compound level, about ten times that of silicon) and glass substrates are also highly expected.
Question 4 | Can Domestic High-Power CW Light Sources Meet Supply Demands Next Year?
On-site conclusion: This question was reconstructed on-site into a better one—categorize first, then answer.
Most domestic NPO solutions have built-in light sources, and 100-130mW is sufficient—domestic supply in this tier is secure.
The real tension lies in the ~400mW tier aligned with NVIDIA's external light source architecture: domestic leader Yuanjie has just made the product and is undergoing customer verification, with batch supply earliest next year; currently, this tier is mainly supplied by Lumentum and Coherent, while Japan's Sumitomo faces supply obstacles due to its technical path (on-site consensus). A side evidence: during the most intense dispute between NPO and CPO, orders for 400mW high-power light sources from Lumentum and Coherent did not drop but rose—indicating that total demand is exploding, it's just that the outside world cannot distinguish the proportion of each. Overseas manufacturers are prioritizing moving wafer capacity to high-power and EML, and low-power old products may be strategically abandoned.
The real question for next year is not "whether supply can meet demand," but: can Yuanjie replicate its script from two years ago and introduce high-power products into the North American major customer system?
Question 5 | Is the Repairability of CPO an Engineering Issue or an Architectural Deadlock?
On-site conclusion: The question with the highest consensus—it is an engineering issue, with yield rate at its core.
NVIDIA's "external light source + replaceable optical components" architecture has already figured out the repair interface—"A lot of design has already been done architecturally, what remains is mainly the yield issue." Currently, the lowest visibility is the yield ramp-up of the NVIDIA and TSMC combination—side information shows it is gradually improving (on-site consensus).
The site provided a set of yield coordinates (all on-site consensus): pluggable is about 95% or above; NPO is currently less than 60%, expected to climb to 80-90% next year; single-channel lasers have broken through 99%—so the 96-channel solution is already feasible; TSMC's CPO process yield is still at a low level.
Every route dispute will ultimately be decided on the yield curve—this is also the reason why the coupling, bonding, and testing equipment segments are generally optimistic.
Question 6 | Who Pays for Silicon Photonics—Chipmakers or Cloud Providers?
On-site conclusion: The overwhelming answer is cloud providers—"whoever decides, pays."
Superficial counterexamples were resolved one by one: NVIDIA's $2 billion investment each in Coherent and Lumentum was interpreted as a means to ensure delivery, lock in capacity, and build exclusivity, rather than truly paying; Meta paid Corning $6 billion to bypass middlemen and directly purchase glass; Google's long-term self-development of TPU and OCS optical switching—every party in the chain is fighting for the "right to pay upstream."
A figure defined this struggle: optical interconnects currently account for about 5-6% of computing center costs (storage and copper-clad laminates are both rising in price, while optical modules haven't), and are expected to exceed 10% after scale-up opticalization. The higher the proportion, the more the paying party cares—this is the economic base of the "who pays" dispute.
The most valuable minority framework is the "Global-China divergence": globally, the three system platforms (NVIDIA's GPU architecture + Broadcom's switching ASIC + TSMC's advanced packaging) have strong bargaining power, and cloud providers pay; domestically, due to conservative procurement logic (prioritizing lower power consumption and lower costs), value distribution will instead moderately flow back to chipmakers. The same question, two markets, two answers.
There is also a dynamic perspective: if the "token factory" business model rises and model companies bypass cloud providers, the payment structure will change again—"Ultimately, it's a game of power, seeing who controls the exit of tokens."
05. Three Cognitions More Important Than the Six Answers
In four and a half hours, the discussion of the six questions repeatedly hit the same three walls. They are more important than the answer to any single question.
First, the Pricing System of the Supply Chain is Shifting: From "Consumer Logic" to "Investment Logic."
Over the past few decades, semiconductors followed consumer logic—no matter the performance, the next generation must be cheaper. AI has brought it into investment logic: hardware makes money directly, and delivery capability and supply stability replace price as the primary competitiveness of the supply chain. On-site evidence: NVIDIA was once willing to "double the price for double the supply" but couldn't get it; a certain quantitative client spent 300 million RMB just to make a hundred chips—"Being one step ahead, you can earn the money back in a week."
Behind this judgment is a complete set of accounts. It was laid out step by step on-site, worth copying down entirely:
Step 1, the per-GW economic account. Taking Meta, Microsoft, and Google as samples: each GW of data center investment is about $25-45 billion; depreciated over five years, the annual cost is about $5-9 billion; while each GW can bring in about $10-15 billion in revenue annually—revenue systematically exceeds depreciation costs for the first time. "In 2026, AI starts making money."
Step 2, GDP anchoring. Global GDP is about $126.3 trillion; if AI drives a 10% growth, it's an incremental output of about $12.6 trillion annually; converted according to the long-term inference gross margin of about 75% for top large models—spend $1, earn $4—the reverse calculation exactly supports about $3 trillion in annual capital expenditure. Jensen Huang's quote "global capital expenditure will reach $3 trillion by 2030" thus balances the books, "not groundless."
Step 3, supply ceiling. The hard constraint is ASML lithography machines: existing stock of over 200 units, adding about 66-70 units annually, projecting to about 200+ GW of new data centers supported annually by 2030; multiplied by the ~$25 billion investment per GW, we get the physical ceiling for the industry's capital expenditure—about $4-5 trillion.
Step 4, who pays. Currently, the money comes almost entirely from several major North American cloud providers (totaling about $1 trillion this year), and they have started borrowing—Amazon has already borrowed tens of billions of dollars. The gap from $1 trillion to $4-5 trillion is huge, and AI will eventually become universal like building bridges and roads—national-level capital from the US, China, and the Middle East may become the most important investors in the next stage.
Step 5, the projection method for long-term demand is also worth noting. The current revenue curve of model companies is essentially still a "single coding curve"—coding is the first monetization scenario to run through. The ceiling of this curve is calculated like this: global IT expenditure plus programmers' wages total about $5 trillion; if 40% of the workload is replaced by AI, it's an annual market space of about $2 trillion. Compared to the current situation—OpenAI and Anthropic's annualized revenue has just entered the tens of billions of dollars level, still two orders of magnitude away from $2 trillion—so "this year is the 0 to 1 of application monetization, next year and the year after are 1 to 10." And this is just the first curve: a year ago, no one expected the explosion of agentic coding, and new monetization curves will continue to emerge in the future—"Time is on the side of the bulls, don't set too many limits for yourself." Further ahead is physical AI: storage usage in the robot era could reach 1,000 times the current level (Micron's data).
The endpoint of the shift in investment logic is one sentence: in the future, the competition is not about how much each token costs, but whose tokens are faster.
Second, Integration and Decoupling: The Route Dispute is Essentially a Confrontation Between Two Industrial Organization Models.
The essence of CPO is "integration of all"—it requires a dominant player that simultaneously controls chips, packaging, and optics; globally, there are only one or two such players like NVIDIA + TSMC; the essence of NPO is "decoupling"—each segment is independently replaceable, which exactly matches China's endowment of "strong modules, scattered ecosystem" and the cloud providers' procurement philosophy of "de-binding." The US is good at industrial integration, while China is good at division of labor. From this, it can be inferred that the real starting signal for domestic CPO is not the yield figure, but the formation of a domestic "integrator" (own AI chips + packaging integration capability).
Third, Technical Optimality Often Loses to Ecosystem Power.
Google's optical module suppliers have historically never exceeded 5, and even in cycles of oversupply, new players are almost never introduced; top manufacturers and second-tier manufacturers purchase the same chips, with a price difference of about 30%; Intel's silicon photonics, which had the most advanced technology ten years ago, almost died before dawn. The same structure repeatedly appeared in the three discussions: "qualification to sit at the table" (binding, certification, positioning) precedes parameter leadership. For all entrants, solve the qualification issue first, then talk about the pros and cons.
01. Numbers Worth Noting
(All are expressions by on-site experts, not independently verified)
Global AI capital expenditure: about $700-950 billion this year, expected $1-1.3 trillion next year; the hard constraint is ASML lithography machine new capacity, projecting a 2030 ceiling of about $4-5 trillion.
Profitability inflection point: each GW data center investment is about $25-45 billion, five-year depreciation annual cost $5-9 billion, annual revenue about $10-15 billion—multiple sets of calculations jointly point to "AI starts making money in 2026."
GDP anchoring: Global GDP about $126.3 trillion × AI driving 10% = $12.6 trillion annual increment; reverse calculated according to top models' ~75% inference gross margin (spend 1 earn 4), supporting about $3 trillion annual capital expenditure.
Demand: 800G about 40 million units this year, 1.6T about 25 million units; combined 150-200 million units next year; NPO and pluggable are not zero-sum, packaged demand growth rate is about 50%.
Yield: Pluggable ≥95% | NPO <60% (80-90% next year) | Single-channel laser >99% | CPO still at a low level.
Shortage: Silicon photonics wafers and EML gap of about 30-40%, at least until next year; "the bottleneck changes every two quarters."
Landscape: Hundreds of companies making 800G, but there are only about 10 A-share optical module companies, of which about a third are not making money (for comparison: 80 listed LED companies, 40 PCB companies); optical interconnects account for about 5-6% of computing center costs, expected to exceed 10% after scale-up opticalization.
Supply projection: Tower's visible capacity by the end of 2027 corresponds to about 80-100 million 800G equivalents (including EML limit of about 200 million); its 2028 new factory is "basically built entirely for NPO."
Material economics: TFLN 8-inch wafer ultimate material cost is about 2,000 RMB, entire line investment about 100 million RMB; strained germanium substrate hole mobility is about ten times that of silicon.
Talent: About 20,000 ultra-precision manufacturing talents globally, and related technologies are no longer being transferred to China—joint cultivation by industry, academia, and research is a long-cycle rigid demand.
07. Pre-Meeting Judgments vs. On-Site Trends: A Comparison Table
Before the meeting, participants submitted written positions on the six topics (25 people showed their cards); after four and a half hours of face-to-face confrontation, the "answers" to most questions shifted—this is exactly the value of closed-door collisions:
08. Beyond the Agenda: Frontline Directions Named
Glass substrates and TGV (universities have built 8-inch lines, NVIDIA executives recently visited TFC Optical to discuss the "glass bridge," on-site consensus); TFLN (low threshold, high performance); new strained germanium substrates; 1100nm silicon-based detectors (the limit wavelength for silicon transparency, CMOS compatible); OCS optical switching (the domestication of core MEMS micromirror arrays was named as a blank); silicon capacitors and silicon-based interposers; multi-wavelength quantum dot light sources; ultrafast laser precision processing (FAU 3D structuring, glass drilling, high-precision micro-grooves); MOCVD epitaxy equipment (overseas manufacturers "simply have no delivery time"); and a cross-boundary brainstorm—"a humanoid robot might need dozens of short-reach optical interconnect units."
09. Golden Quotes from the Site
(Anonymously paraphrased)
"What we fear is not oversupply, but that it becomes obsolete just as we finish making it."
"Whoever the AI chip chooses is the only one qualified to do CPO with them; otherwise, you don't even have the qualification to sit at the table."
"Moore's Law fails in the manufacturing of optics—this is precisely the most friendly place for domestic players."
"Without manufacturing capability, you cannot hone design capability—chips are made through iteration."
"Regarding the NPO socket, everyone thinks you just plug it in and it works, but it's actually not."
"Customers don't care if the price is halved or even cut to one-tenth; in the future, the competition is not about how much each token costs, but whose tokens are faster."
"The optical communication supply chain is facing a reconstruction of value distribution—digging to the first principle, it's about who holds the premium pricing power."
"Putting light into the chip is what silicon photonics is about—this is the core to solving AI computing power."
"A change in materials will bring about a change in the industry."
"Too early never works—the silicon photonics leader ten years ago almost died later."
"Time is on the side of the bulls—we are at the beginning of a great revolution, don't set too many limits for yourself."
"Traffic is in the lively places, value lies in the quiet places."
10. Epilogue: Divergences Remain, Answers to be Checked in October
This meeting did not attempt to eliminate divergences—divergences themselves are the most real state of the industry right now. In four and a half hours, only one of the six questions reached a nearly consistent answer (repairability is an engineering issue), while the other five left live opposites.
But several things reached an action consensus before leaving: establish China's own silicon photonics technology roadmap (referencing overseas institutions, dynamically updated, there are already initiators willing to take the lead, which will become the first task of the Silicon Photonics Special Committee of the "Tech Research 100" being prepared); appeal to form a unified CPO route suitable for China; promote CPO industry-wide process collaboration; and cross-enterprise joint research on the overall "electronics + optics + packaging" solution.
The organizer promised to review on a quarterly basis: in three months, every judgment made today will be brought out and checked item by item—those who are right are prophets, and those who are wrong will review together.
Next session, October.