:
1. Photonics TAM goes from $14B -> $154B In just two years time, and it's likely going to keep scaling past 2030 as it's the next generation architecture of choice.
It's not going away in 1 year. It's not going away in 3 years, which is why
$LITE
premiums keep going higher since they're backlogged into 2028.
This isn't a "trade", it's the core chokepoint and IP holder for the next generation of photonics.
And it's a comfortable hold for the next few years as they scale to become the next
The risk I personally see (since they're already qualified with so many players), it's mainly how much TAM they can capture of the overall optical supercycle. (And potential risks with Win Semi volume ramp, but Win is massive so I can sleep tightly there).
As just supplying lasers isn't enough to justify valuation.
It's TAM expansion downward into making the entire ELS or entire pluggable transceiver that makes these laser companies so valuable.
Then afterward, they can vertically integrating upward for gross margin expansion upward like
- Is the purest exposure, without the messy financials of SKC Absolics, as the next advanced packaging shift for glass substrates.
Almost every single major semi company from
, where if there's more foundry capex, ASML scales up. But if there's downturns, these tend to perform poorly, and don't capture all the volume ramp that happens after.
However, if the MC is $650m and they're making $100-200M, revenue per costumer volume ramped, the amount they make from the glass substrate cycle will likely exceed current valuations.
And they'll have baseline fundamentals (as more companies adopt the packaging shift), that keeps their valuation up.
It's just a waiting game for volume ramp at this point.
3.
but for America + Photonics. It's like saying Intel is not a long term investment.
Guess where all your optical transcivers are made?
China. Thailand. Malaysia. If you look at Innolight, Eoptolink,
, and others.
AOI is building the largest Made in America supply chains for both CW laser fab, as well as 800g, 1.6T assembly.
Yes, there are pluggable cycle ups and downs to this as well. There's going to be a wave for 1.6T next year, then CPO cannibalizes pluggables down the road.
But since they make the entire supply chain in house, they have extreme optionality for other segments. And like
older gen-GPUs, there's going to be sovereign DC requirements for older gen pluggables from names like
$AAOI.
It's likely going to keep rising as it hits that $400m+/month revenue target H2 2026.
There's just a lot of different short term volatility along the way like the $600m dilution.
4.
The world is currently bottlenecked both on the epiwafer level from Landmark comments and InP substrate levels.
Their financials were track but the raw book value, and value they hold to the entire Western supply chain... completely justifies their valuation. And other optical companies will not let their core upstream supply chain go under.
As these tens of millions worth of materials would screw up tens of billions worth of downstream products.
Again photonics is the next generation architecture required to scale AI. It's not Quantum where it's just "In development".
It's literally here and the architecture of choice by
$NVDA.
I would not be surprised if all of these are a lot higher in 3-4 years time.
People who think it's one and done in 3 months time "only because I mentioned it" don't know what they're talking about.
Institutions would have bought up the name eventually (like Point 72 on
) and retail would only find out after their valuations are 600% higher.
Should really do the research before adding comments like these:
These are all forward growth companies that require in-depth supply chain knowledge.
Most people are focused on silicon.
The real bottleneck nobody’s talking about? Indium Phosphide.
$NVDA needs faster chips.
Faster chips need faster interconnects.
Faster interconnects need InP-based lasers. And InP supply is critically constrained.
WHAT IS InP AND WHY DOES IT MATTER?
Indium Phosphide is the substrate behind high-speed optical components — the lasers and photodetectors moving data at 800G and 1.6T speeds inside AI data centers. Silicon simply can’t do what InP does at these speeds.
No InP = No optical interconnects = AI infrastructure hits a wall.
THE FULL InP VALUE CHAIN
𝗦𝘂𝗯𝘀𝘁𝗿𝗮𝘁𝗲𝘀 & 𝗪𝗮𝗳𝗲𝗿𝘀
→ $AXTI — one of the only publicly traded InP substrate suppliers in the West
→ $IQEPF — epitaxial wafer supplier feeding InP laser production
𝗖𝗼𝗺𝗽𝗼𝗻𝗲𝗻𝘁𝘀 & 𝗟𝗮𝘀𝗲𝗿𝘀
→ $COHR — vertically integrated, owns InP laser fabs
→ $IPGP — fiber and InP laser exposure
→ $AAOI — transceiver/laser components
→ $LITE — InP laser supplier to hyperscalers
𝗦𝗶𝗹𝗶𝗰𝗼𝗻 𝗣𝗵𝗼𝘁𝗼𝗻𝗶𝗰𝘀 / 𝗙𝗼𝘂𝗻𝗱𝗿𝗶𝗲𝘀
→ $TSEM — InP photonics foundry capabilities
→ $GFS — compound semiconductor exposure
𝗣𝘂𝗿𝗲-𝗣𝗹𝗮𝘆 𝗦𝗽𝗲𝗰𝘂𝗹𝗮𝘁𝗶𝘃𝗲
→ $POET — optical interposer platform designed to integrate InP lasers at scale, potentially solving the bottleneck directly
→ $LWLG — electro-optic polymer platform, InP alternative play
→ $ALMU / $SIVEF — smaller speculative names in the photonics supply chain
𝗧𝗲𝘀𝘁𝗶𝗻𝗴 & 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻
→ $AEHR — wafer-level burn-in testing for photonic chips
→ $KEYS — optical and high-speed signal testing
WHY SUPPLY CAN’T JUST SCALE OVERNIGHT
InP isn’t silicon.
→ Only a handful of facilities globally can produce InP substrates
→ Building new InP fabs takes years and billions
→ Geopolitical concentration risk is real — much of the supply chain runs through Asia
→ Hyperscaler demand for 800G/1.6T is accelerating faster than supply can respond
This is a structural bottleneck — not a temporary one.
THE INVESTMENT THESIS IN ONE LINE
The AI buildout runs through InP.
And InP supply is stuck.
The picks-and-shovels play isn’t just $NVDA.
It might be the substrate nobody’s heard of yet.
Not financial advice.
HOW DOES $POET ($3.14B) HAVE A HIGHER VALUATION THAN FOCI (3363, $3.1B)???
FOCI IS LITERALLY THE BOTTLENECK FOR CPO VOLUME RAMP AND MAIN SUPPLIER FOR $TSM AND $NVDA.
High conviction Foci outperforms once institutions find this name.
Also, can Foci management please pursue NASDAQ ADR like $HIMX? Thank you.
So here's the napkin math I did on Nextronics (8147) when I went long.
They're the $NVDA CPO supplier for CPO connectors and cage thermal modules.
And I modeled around 2 FWD p/e for 2028, which is why I think risk-reward is very compelling for a potential 10x rerating to ~$2B+ MC in 2028.
Just for their CPO exposure:
-> CPO connector runs roughly $15 to $25
-> ELS thermal cages, maybe ~$50 from est.
18 units per switch: 18x50 = ~$900
CPO Connectors: 72 Optical Engines per switch 72 x $15 = $1,080
(If $NVDA scales their Spectrum-X switch, it goes to $1,920 for CPO connectors).
Total Nextronics Content: ~$1,980 (rounded to $2k for calculations) in conservative case.
Implied BOM % of rack: 0.08%. Maybe ~1.5% of switch.
This looks microscopic to institutions so it probably is ignored.
Is it material to Nextronics, a ~$200m company?
Yes, absolute massive.
For calculations: Applying 50% haircut to Nextronics' share of the Nvidia connector market/cage market because of multi-source.
And I’m using GS projections, and assuming $AVGO, $MRVL, ASIC CPO ecosystem is 30% size of $NVDA.
Net Income Margin: 22.4% (at 38% GM)- 24.0% (at 40% GM).
But going off other projections from just, a rack shipments:
2026: CPO revenue ~10.1M, net income (22.4%)
~2.26M + $12.5M base = $14.7M (540k units for connectors, cage, 40K units, already divided by 50%)
2027: CPO revenue: ~$172M, net income (22.4%): ~$38.53M = $51.03M (~8M units for connectors, ~1.03M units for cage)
2028 scale up expansion: CPO revenue: $450m, net income: $100.93M, ~$11.3M base (~40M units for connectors, 2.98M unit for cages, eg. Nvidia ELS volume is 19.9M)
So implied fwd p/e 15.4x for 2026, 4.45x for 2027, 2x for 2028.
Of course at scale, blended margins might go down, there might be other players bringing market share down to like 25%, etc. and projections might be more or less than GS.
But regardless seems highly asymmetrical even if I'm off by a whole 50%.
2028 is usually the massive re-rating for CPO players, 2026 is still really early.
Hope my math is right, but 20x fwd p/e multiple would be $2.26B MC.
Even if we drop:
-> market share to just 15%.
-> compress their net income margin down to 14%.
-> connector ASP to $10.
At a 20x multiple, the stock would still achieve a ~4.5x return to a $1B+ market cap.
We'll see if this is right or not. (NFA, just speculative financial modeling)
THE IPO OF 2026 IS THURSDAY
$CBRS — Cerebras Systems hits Nasdaq on May 14th
And this isn’t just another AI hype listing. This is the company that built a chip the size of a dinner plate to go to war with $NVDA.
THE TECHNOLOGY
→ Wafer-Scale Engine (WSE) — 58x larger than Nvidia’s B200, with 19x more transistors, 250x more on-chip memory, and 2,625x more memory bandwidth 
→ The entire silicon wafer IS the chip — no inter-chip communication latency
→ Purpose-built for AI inference, not training — where the next trillion-dollar battle is being fought
THE DEAL TERMS (UPDATED)
→ IPO price range raised to $150–$160/share from the original $115–$125 range 
→ Shares offered upsized to 30M from 28M — raising up to ~$4.8B 
→ Book is oversubscribed 20x — demand is historic 
→ Largest IPO globally so far in 2026 
THE FUNDAMENTALS
→ 2025 revenue: ~$510M, up from ~$290M in 2024 
→ OpenAI signed a $20B+ deal for 750MW of Cerebras-backed compute through 2028 
→ AWS deploying CS-3 systems in its data centers and distributing through Amazon Bedrock 
→ Implied valuation at top of new range: ~$32B+
THE RISKS
→ Two customers account for 86% of revenue — concentration risk is real 
→ GAAP operating losses persist despite strong top-line growth
→ Previous IPO was pulled in 2024 over CFIUS/G42 (UAE) scrutiny — now cleared, but the overhang is part of the history
→ 20x oversubscribed = first-day pop priced in, early volatility likely
THE BIGGER PICTURE
Cerebras isn’t trying to beat $NVDA at training. It’s attacking the inference bottleneck — the part of AI compute that scales with every user query, every agent, every real-time application. If agentic AI is the next wave, low-latency inference silicon becomes the most strategic asset in the stack.
Just reiterating my disbelief:
I have never seen a sector more bullish than CPO.
GS reported Optical TAM 9X from $15b in 2026 -> US$154b in 2028
CPO making up $91B of that.
Starting from ~$164M (Modor for 2026 / sampling) to $91 Billion (GS 2028)
55,000%+ CPO growth curve starting from today
This is exactly why algorithms / analysts mess up because they might look at TTM revenue at these CPO names.
But everything happens in the next two years with $SIVE to Shunsin to MSSCorps to $SOI.
This is Zero to 100 from a massive architectural shift pushed by $NVDA.
I genuinely still don't think retail or markets understand what's coming yet.
4. Tell Claude to argue against you.
Most people use AI to confirm their bias. Flip it.
"I'm long $NVDA. Give me the bear case a short seller would make."
You'll find blind spots you didn't know you had. I've killed trades because of this. Saved me real money.
1. Nebius ~ $NBIS
They’ve captured a $50B approaching backlog for (2027-2031) consisting of…
- $27B contract from $META for AI compute capacity.
-$19B GPU compute deal from $MSFT.
- $2B investment by $NVDA.
This has helped Nebius sustain 546% YoY revenue growth, & grow a $4.6B cash pile.
It’s clear this name will continue to pile contracts and expand exponentially.
Targets include $220 before Q3 26’ and then $330 before Q2 27’.
ROBOTICS: The next decade won’t be won by the best demo. It’ll be won by whoever solves deployment at scale.
Everyone sees humanoid demos. Few see the bottlenecks.
The real opportunity in robotics isn’t just the robot — it’s the infrastructure behind it.
Biggest chokepoints:
Actuators & precision mechanics
Joint systems are expensive, complex, and supply-constrained. Every robot depends on them.
Battery life
Most humanoids still can’t work a full shift. Energy density remains a major hurdle.
Dexterous hands
Walking is improving fast. Manipulation is still the real unlock.
Training data
LLMs had the internet. Robots need physical-world action data. That gap is massive.
Software stack
Deployment complexity is still too high. Robotics needs its “Python moment.”
Unit economics
Industrial customers need reliability first, not cool demos.
Rare earth supply chain
Magnets, batteries, and precision parts are the hidden chokepoints.
My view:
The winners may not just be platform builders like $TSLA or AI leaders like $NVDA.
The bigger compounding opportunity could be in picks-and-shovels:
Robotics is not one market.
Medical is already scaled.
Hospitality is scaling now.
Industrial is crossing the gap.
Humanoids are still early.
The demo problem is mostly solved.
The deployment problem is where the money will be made.
$POET and $LWLG — The Photonics Long Game
Two names. Same mega-theme. Very different risk profiles. Here’s why both deserve a spot on your watchlist 👇
The Tailwind (Same for Both)
→ Optical transceiver market doubling from $5B → $10B by 2026
→ Projected to hit $100B by 2030
→ AI data centers NEED faster, cheaper, lower-power interconnects
→ Silicon photonics is the bottleneck — these companies are the solution
$POET — The Commercial Ramp Play
→ Real revenue: $5M+ production order secured ✅
→ 30,000+ optical engines shipping in 2026
→ 800G → 1.6T → 3.2T product roadmap
→ $430M cash, no debt — serious runway
→ Partners: Foxconn, Luxshare, Mitsubishi Electric
→ Collaborating with $QUBT on 3.2Tbps TFLN engines
$LWLG — The IP Platform Bet
→ Proprietary EO polymer tech — higher speed, lower power, smaller form factor
→ 4 Fortune Global 500 companies now in Stage 3 (prototype → product)
→ Targeting 200G and 400G per lane for hyperscale AI factories
→ Foundry access expanding: Tower Semiconductor + GlobalFoundries PDK integration
→ If EO polymer becomes the modulator standard → royalty/licensing goldmine
$POET = more de-risked, real orders, manufacturing underway
$LWLG = pure asymmetric optionality — if the polymer wins, it wins big
Best combo.
Photonics is the infrastructure layer most investors are still sleeping on.
The AI trade isn’t just $NVDA.
It runs on light.
Not financial advice.
The AI Infrastructure Bottleneck
AI demand isn’t slowing. It’s colliding with reality. Power costs are rising. Communities are pushing back. Permits are getting harder. Project cancellations are climbing. This isn’t weak AI demand. This is infrastructure hitting its ceiling. Compute scaled in 18 months what normally takes a decade.
Data centers went from background infrastructure to the most power-hungry real estate on earth — almost overnight. Utilities weren’t ready. Grids weren’t ready. Permitting systems built for a slower world weren’t ready.
So now you have the most capital-rich companies in history competing for a fixed, constrained resource: Megawatts.
That collision created a new capital cycle:
AI → Energy → Infrastructure
The market hasn’t fully priced the downstream implications yet. When datacenter capacity gets delayed or cancelled, the reflex is to ask:
“Is AI demand softening?”
Wrong question.
The right question is: “Who benefits when supply tightens?”
Less derisked capacity doesn’t reduce demand. It concentrates it.
And concentrated demand is a pricing event.
When power supply tightens:
→ Frontier labs fight for compute at premium pricing
→ Hyperscalers delay expansion timelines, reshuffling capex
→ Chip, memory, ASIC, and CPU ecosystems feel the ripple — not because demand died, but because deployment slowed
→ Whoever controls reliable, permitted, grid-connected power becomes the most valuable asset in the AI stack
This is the Silicon-to-Substation Shift.
Just like bandwidth defined the internet era — whoever controlled the pipes controlled the economy — reliable power generation is becoming the defining constraint of the AI economy.
The bottleneck has moved up the stack. The market always chases the obvious trade first. $NVDA.$AMD. Memory. Networking. The picks and shovels everyone already knows.
But when the bottleneck shifts — the alpha moves too. Right now, the bottleneck isn’t silicon.
It’s megawatts, transformers, transmission lines, and building permits.
The second-order trade is where the real opportunity lives — across five infrastructure layers most investors haven’t rotated into yet:
Power Generation
Nuclear is back. Natural gas is back. Utilities signing 20-year offtake agreements with hyperscalers.
The cleanest, most reliable, 24/7 baseload power is suddenly the scarcest commodity in tech.
$VST → baseload + nuclear, direct hyperscaler relationships
$CEG → largest nuclear operator in the US, massive AI offtake pipeline
$SMR / $OKLO → next-gen modular nuclear, secular long-term positioning
Grid Infrastructure
Transformers. Substations. Transmission lines.
The hardware between the power plant and the server rack is now a national bottleneck — with multi-year lead times.
$PWR → grid construction & transmission at scale
$ETN → electrical components backbone, transformer exposure, margin expansion
Energy Storage
Intermittent renewables can’t power AI campuses alone.
Grid-scale battery storage and microgrid systems bridge the gap — especially as utilities demand more dispatchable flexibility before approving new load connections.
$FLNC → utility-scale storage, growing backlog, grid stabilization play
Cooling Infrastructure
A 100MW AI cluster generates extraordinary heat density.
Air cooling is dead at scale. Liquid cooling, immersion systems, and thermal management are now standard spec in every new hyperscale build — one of the most underappreciated sub-themes in the entire stack.
Datacenter Construction & Power Systems
Before a single GPU goes online — somebody pours the concrete, runs the conduit, installs the switchgear, and commissions the facility.
Construction and power systems backlogs are extending years out.
$STRL → datacenter construction, revenue inflecting hard
$BW → industrial power systems, 1.2GW AI datacenter deployment
The thesis is simple:
AI isn’t running out of demand.
It’s running into megawatts.
Not financial advice.
@daniel_koss thanks! small positions on $NVTS though. I find photonics theme more compelling overall.
People don't need to get exposure to every story that pops up with $NVDA.
$NASA — The Space ETF
Tema just launched the first pure-play space ETF with actual SpaceX exposure.
Here’s what’s inside-
Top Holdings:
$ASTS — 7.91%
$RKLB — 6.95%
$PL — 6.68%
$FLY — 5.42%
$LUNR — 4.63%
$SATS — 3.92%
+ SpaceX (private SPV) ~10–15%
Key Stats:
AUM: ~$144M
Expense ratio: 0.75%
20–40 holdings
Launched: March 31, 2026
Up ~17–20% since debut
Why it matters:
A space ETF without SpaceX is like a semis ETF without $NVDA. NASA fixes that. SpaceX IPO expected mid-2026
Not financial advice. DYOR.
Most investors are always chasing the last wave.
The AI infrastructure cycle is playing out in layers — and each one is bigger than the last.
Here’s the full roadmap:
Wave 1: Semiconductors ✅ (Priced In)
$NVDA.$AMD.$AVGO.
The GPU arms race. Everyone knows this trade.
Raw compute was the foundation.
Most of the upside? Already captured.
Wave 2: Memory & Storage ✅ (Confirming)
More data. More inference. More throughput.
Flash, DRAM, HBM all followed.
$SNDK ’s recent move is the latest confirmation.
The storage bottleneck is real — and the market is finally catching up.
Wave 3: Photonics & Optical Networking 👀 (In Motion NOW)
This is what most retail investors are still sleeping on. You cannot move AI workloads at scale with copper wire.
Data centers need ultra-fast, low-latency interconnects — and photonics is the only answer.
Silicon photonics is actively replacing copper-based communication inside and between hyperscaler campuses.
The plays institutions are quietly accumulating:
$AAOI — pure-play optical components, deep data center exposure
$COHR — vertically integrated photonics at scale
$LITE — transceivers powering the hyperscaler build-out
$POET — integrated optical engines, early-stage, asymmetric upside
$CIEN — backbone optical networking infrastructure
$FN — precision optical manufacturing, best-in-class margins
This wave is not coming. It is here.
Wave 4: Power & Energy Infrastructure (Early Stage)
AI doesn’t run on ambition. It runs on electricity. Data centers are projected to consume 8–10% of U.S. power by 2030. That is not a tailwind.
That is a structural demand shock the grid is not prepared for.
Nuclear is the only baseload solution that can sit next to a hyperscaler campus and deliver consistent, carbon-light power at scale.
The plays:
$CEG & $VST — utilities with nuclear-powered AI offtake contracts
$CCJ — uranium supply, the fuel behind the revival
$OKLO & $SMR — next-gen small modular reactors
$VRT — liquid cooling and power management inside data centers
The capex commitments from $MSFT, $AMZN , and $GOOG are already locked in. The power infrastructure just needs to catch up. This wave is early. That is the opportunity.
Wave 5: Robotics 🤖 (Loading…)
Once AI is trained, powered, and connected — it needs a body.
Physical deployment into warehouses, factories, logistics, and healthcare.
This is the final and potentially the largest layer of the entire cycle.
Hardware-software integration is still 12–24 months from full ignition.
But the smart money is already laying the groundwork quietly.
The pattern is simple:
Each layer enables the one after it.
→ Semis built the brain
→ Memory gave it recall
→ Photonics gave it nerves
→ Power gives it fuel
→ Robotics gives it hands
The investors winning this decade are not chasing yesterday’s leader.
They are identifying tomorrow’s bottleneck — and positioning before the crowd arrives.
The question is not which wave already ran.
The question is: which wave are YOU in front of?
$AAOI$COHR$LITE$POET$CIEN$FN$CEG$VST$OKLO$SMR$CCJ$VRT
DOYR , bookmark this and retweet for others.
$NBIS has $46 billion in contracted revenue from $MSFT and $META.
> $2 billion investment from $NVDA.
> 310MW breaking ground in Finland.
> A Missouri GW campus just approved.
> Revenue guided $91M to +$3B EOY 26.
That is EXECUTED backlog.
The stock is up 400% in twelve months and still trading at a discount. The market has not priced in the full contract stack. It is STILL pricing in doubt about execution.
I understand that doubt. The capital plan is aggressive. The buildout is enormous.
CEO Arkadiy Volozh has to deliver 800MW to 1GW of connected capacity by year-end to stay on track.
But the risk here is execution pace now, not demand. $META and $MSFT are not letters of intent.
They are binding multi-year commitments from two of the three largest technology companies on earth.
When hyperscalers write $46 billion in contracts to one neocloud, they are not doing due diligence on a wish list.
They are securing infrastructure they cannot build fast enough themselves.
Demand in this space is STILL insane.
-BP
Not financial advice. Do your own due diligence.
My $MRVL position is already up 35%.
I think they are one of the safest plays into photonics. Financially very healthy and they are using their cash do to some great acquisition.
Its Celestial AI deal adds Photonic Fabric, strengthening its position in scale-up connectivity for next-gen data centers.
They also have a strategic partnership with $NVDA. They are working together on NVLink Fusion, custom silicon, scale-up networking, and silicon photonics so customers can build AI systems that stay compatible with NVIDIA’s ecosystem.
NVIDIA announced a $2 billion equity investment in Marvell as part of the expanded partnership.
NVIDIA gets a broader, more flexible AI hardware stack, and Marvell gets deeper access to NVIDIA’s AI factory and AI-RAN ecosystem.
With the whole focus on photonics companies like $LITE, $COHR, and $AAOI, it surprises me that a company like $FN is completely ignored by X investors.
Fabrinet is essentially the manufacturing backbone for the entire optical industry. Companies like Lumentum and Coherent pour billions into R&D to design cutting-edge lasers and transceivers, but they outsource the actual building, testing, and packaging to $FN. This allows them to scale without spending billions on their own factories.
Furthermore, $FN is already manufacturing optical cables directly for $NVDA. Nvidia designs its own proprietary Active Optical Cables (AOCs) to connect its GPUs within server racks. Instead of buying these from a middleman, Nvidia goes straight to Fabrinet to build them.
AI data centers are entirely dependent on moving data between thousands of GPUs instantly. The bottleneck is optical networking. Because Fabrinet manufactures these ultra-precise, high-margin optical components, they hold a highly specialized and lucrative position.
The ceiling on $FN might be lower than the volatile designers like $LITE and $COHR, but for investors who want a safer route into photonics, $FN is perfectly positioned. We know photonics is a bottleneck, but $FN is the bottleneck within the bottleneck.
Note: I'm not an investor in $FN, need to do a bit more research before making my decision.
$MRVL up almost 11% in pre-earnings after they announced a strategic partnership with $NVDA. This includes a direct $2 billion investment from the chip giant.
Quick breakdown
AI Factory Integration: The partnership connects Marvell directly to NVIDIA’s AI factory and AI-RAN ecosystem through the NVIDIA NVLink Fusion rack-scale platform.
Custom Silicon & Networking: Under the agreement, Marvell will supply custom XPUs and scale-up networking that is fully compatible with NVIDIA's NVLink ecosystem.
NVIDIA's Hardware Contribution: NVIDIA will supply its supporting tech stack—including Vera CPUs, ConnectX NICs, Bluefield DPUs, and Spectrum-X switches—allowing customers to build out semi-custom, heterogeneous AI infrastructure.
Telecommunications Push: The two companies are also collaborating on silicon photonics with the goal of transforming traditional telecom networks into AI infrastructure for 5G and 6G.
$NVDA isn’t just building chips, but rather buying the entire AI stack…
Below are some of Nividia’s most notable investments:
~ $5B in $INTC (Intel)
~ $2B in $CRWV (CoreWeave)
~ $2B in $SNPS (Synopsis)
~ $2B in $LITE (Lumentum)
~ $2B in $COHR (Coherent)
~ $1B in $NOK (Nokia)
~ $30M in $NBIS (Nebius)
Following this list, here are some of the top private companies invested in by $NVDA.
~ $30B in OpenAI
~ $10B in Anthropic
~ $5B in xAI
~ $500M in Wayve
~ $350M in Mistral AI
~ $300M in Crusoe
~ $250M in Scale AI
~ $250M in Cohere
~ $150M in Figure AI
& along with many other smaller sized investments…
$AMZN is currently sitting at decade low valuations, and is set up for a breakout…
Just today $AMZN announced they built an AI chip to compete with $NVDA.
If you missed the move from $150 to $320+ on $GOOGL then don’t miss $AMZN.
$350+ is incoming next year.
Mark my words… https://t.co/ukGHxvlASw
Interesting story by WSJ on Trump-Xi talks in Busan and $NVDA.
- Before meeting Xi, Trump weighed Nvidia CEO Jensen Huang’s push to allow Blackwell AI chip exports to China—a major potential policy reversal.
- Top officials, including Rubio and Lutnick, opposed the move, warning it would endanger U.S. national security while aiding China’s AI ambitions.
- United opposition from Trump's aide is reportedly making Trump decide not to discuss $NVDA chips during the meeting with Xi.
EMC Chairman Ding-Yu Dong: "EMC is currently operating at FULL CAPACITY AND ALL OUTPUT SOLD"
The company launched a three-year expansion plan last year, under which total capacity will increase by about 70% by 2027, with an additional 1 million sheets added in both 2025 and 2026. Output value will grow proportionally.
“This is how it will be for the next few years — full capacity, full sales,” he said, adding that the strong market demand continues and all of EMC’s production lines remain fully loaded.
In terms of product technology, EMC’s current flagship product line remains the M8 series, but Ding-Yu revealed that the next-generation M9 product has COMPLETED CUSTOMER QUALIFICATION and will begin ramping up production in the 2H26.
"EMC's M9 is the first to be qualified and will also be the first to enter mass production — just like our M8,” he emphasized, highlighting the company’s leadership in next-generation high-end materials."
Discussing AI development trends, Ding-Yu stated that data center and server demand will continue to surge, adding,
“It won’t be just $NVDA.” He expects that within three years, both laptops and database servers will move toward full white-label (ODM/OEM) models, with design and production increasingly dominated by ODM/OEM manufacturers.
He summarized EMC’s role in one sentence: “We are like cement — whether it’s a skyscraper or a thatched hut, everyone needs our raw materials.”
KGI: According to our recent check, demand for $TSM CoWoS in 2026 has become even stronger compared with two weeks ago, mainly driven by $NVDA.
We estimate $TSM's CoWoS monthly capacity will expand from 70,000 wafers at the end of 2025 to 115,000 wafers by the end of 2026 (previously estimated at 110,000 wafers).
As we entering into early Q4, ODM/OEM performance might start getting attention again, and unsurprisly on its shipment for $NVDA.
MS raise rack shipment for Wistron from 6.8k to 10k for 2026.
Some other good data points (raise CoWoS guide) from downstream are emerging as well.
While GB300 shipments will begin soon, volumes are expected to become more meaningful in late Q4'25 and Q1’26
