PackagingMemoryPhotonicsIndia📅 Thu 29 Oct🔬 Research radar
Memory makers sell out to 2029, a photonic switch for AI racks, and 5 new chip ideas
SK hynix and Samsung say AI memory is sold out for years. PsiQuantum shows a light-based switch for AI networks. Plus four more new papers and products, each scored.
Step 1 of 3 · Pick of the day: The one story to read first today.
⭐ BREAKTHROUGH OF THE DAY · Thu 29 Oct
PsiQuantum shows a light switch that rewires AI data centres in 3 ms
Stage 2 · Prototype
A quantum-computing company turned its low-loss silicon-photonics process into a switch for today's AI data centres, made at a commercial foundry, with samples already with prospective customers. Lab results💡 are company-reported, so we score it Stage 2.
Step 2First switches in real AI clusters if tests go well.
Step 3Could be a standard part of AI networks, or lose to cheaper electrical switches.
Timeline steps are our judgment unless marked as company plan.
What they found: PsiQuantum is a quantum-computer company. It turned its light-chip process into a light switch for AI networks. The switch has 8 inputs and 8 outputs. The company reports very low loss and little mix-up between signs. It changes routes in under 3 thousandths of a second. It is sharing samples with a few possible buyers.
🧒 In simple words: Computers in AI buildings talk with flashes of light. This chip points the light to the right computer, fast.
🧩 Think of it like: Like a railway junction that swaps tracks in a blink, so data trains of light never have to stop and be turned into electricity.
How close to real use (0–100)46/100
Years to market: 1-4
Our judgment
🎢 Talk vs. real progress
Talk 50
Real progress 46
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
Optical switching is crowded (Google, Lumentum, Coherent and MEMS makers). No known litigation.
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
PsiQuantumGlobalFoundriesgiant cloud companies building AI clusters
Other possible winners: PsiQuantum, GlobalFoundries, giant cloud companies building AI clusters (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
PsiQuantum is a quantum-computer company. It turned its light-chip process into a light switch for AI networks. The switch has 8 inputs and 8 outputs. The company reports very low loss and little mix-up between signs. …
How close to real use (0–100) 46/100, stage 2 on our ladder.
score /100: half of how-ready score 46 + 0 (no big event expected soon)
⛔ Why not yet
Results are company-reported lab results, not independently verified; only an 8×8 port count so far
Google already uses MEMS-based optical switches in production, so the market has a strong incumbent
Timeline to impact: about 1-4 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: A product launch "in coming months", a switch bigger than 8 by 8, and a named data-centre buyer.
A named paying customer or order
Independent replication / outside test results
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Samples tested by a few big cloud firms.
First switches in real AI clusters if tests go well.
Could be a standard part of AI networks, or lose to cheaper electrical switches.
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
PsiQuantum is a quantum-computer company.
Who could win: PsiQuantum, GlobalFoundries, giant cloud companies building AI clusters
🐻 What could go wrong
Results are company-reported lab results, not independently verified; only an 8×8 port count so far source
Google already uses MEMS-based optical switches in production, so the market has a strong incumbent source
All-optical switching has a long history of hype in telecom (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
2/5 — Working prototype chip
55%
13.8
Manufacturability
4/5
20%
16.0
Supply chain
3/5
10%
6.0
Corporate backing
3/5
10%
6.0
Independent groups
4/5
5%
4.0
How close to real use (0–100)
46/100
Talk score = 10×claim boldness (2/5) + 10×attention (3/5) = 50. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: A product launch "in coming months", a switch bigger than 8 by 8, and a named data-centre buyer.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Samples tested by a few big cloud firms.
📋 Executive briefing teaser free
Light switches save power and time in AI buildings.
Light switches save power and time in AI buildings. They skip turning light into electricity and back. Google already uses another kind. PsiQuantum makes a chip version in a normal chip factory. These are the company's own lab results. Nobody else has checked them yet.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Your AI data centre has a traffic problem. Every switch turns light into electricity and back. PsiQuantum just showed a chip that switches the light itself, in 3 milliseconds. Samples are out. Sources: PsiQuantum, IEEE ComSoc.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Step 1More simulation papers; HBM💡 (fast memory stacked next to AI chips) makers test ideas.
Step 2Possible first compute features in custom HBM base dies.
Step 3Common in AI memory, if standards and software follow.
Timeline steps are our judgment unless marked as company plan.
What they found: The authors add small helpers to the bottom layer of fast AI memory (HBM4). The main chip reads the question. The memory helper writes the answer word by word. In a computer model it was up to 2.23 times faster. It used up to 37.9% less energy.
🧒 In simple words: The computer's brain keeps running to its notebook. This idea puts a helper inside the notebook to do some work.
🧩 Think of it like: Instead of running to the library for every sentence, the librarian writes the answer at the shelf and only hands you the result.
How close to real use (0–100)27/100
Years to market: 4-8
Our judgment
🎢 Talk vs. real progress
Talk 40
Real progress 27
Some hype; watch for proof (our judgment)
⚖️ Patent / patents risk
Processing-in-memory is heavily patented by Samsung, SK hynix and others. No known litigation.
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
In plain words: The computer's brain keeps running to its notebook. This idea puts a helper inside the notebook to do some work.
📊 Stocks tied to this
Company
Stock code
How it is linked
Role
How sure we are
No listed company directly tied yet. Developer: Seoul National University (Jae-Joon Kim group). Watch for a listed partner or IPO💡.
Other possible winners: SK hynix, Samsung, Micron (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
The authors add small helpers to the bottom layer of fast AI memory (HBM4). The main chip reads the question. The memory helper writes the answer word by word. In a computer model it was up to 2.23 times faster. It used …
How close to real use (0–100) 27/100, stage 1 on our ladder.
score /100: half of how-ready score 27 + 0 (no big event expected soon) − 8×1 warning flag(s) · no listed stock code linked → ⚪ Too early: still early
⛔ Why not yet
Simulation only, no fabricated chip
Needs HBM vendors and GPU💡 makers to agree on a new interface; earlier in-memory compute products saw limited adoption
Timeline to impact: about 4-8 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Whether a memory maker adds helpers to its memory, and any test on a real chip.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More simulation papers; HBM makers test ideas.
Possible first compute features in custom HBM base dies.
Common in AI memory, if standards and software follow.
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
Needs HBM vendors and GPU makers to agree on a new interface; earlier in-memory compute products saw limited adoption (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
0/5 — Paper / simulation only
55%
0.0
Manufacturability
4/5
20%
16.0
Supply chain
3/5
10%
6.0
Corporate backing
1/5
10%
2.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
27/100
Talk score = 10×claim boldness (3/5) + 10×attention (1/5) = 40. Risk of a fight over who owns the idea: Medium (crowding 4/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Whether a memory maker adds helpers to its memory, and any test on a real chip.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More simulation papers; HBM makers test ideas.
📋 Executive briefing teaser free
New AI memory has room for small helpers in its bottom layer.
New AI memory has room for small helpers in its bottom layer. This design splits the work between the chip and the memory. Gains are only in a computer model. Memory makers would need to agree.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
AI answers are slow because data shuttles back and forth from memory. A Seoul team put tiny processors inside the memory stack. Simulated result: up to 2.2× faster answers. Real chips needed. Source: arXiv.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Timeline steps are our judgment unless marked as company plan.
What they found: Thin films of a material called bismuth ferrite usually stop working near the surface. That is called a "dead layer". The team removed it. The material still worked at about 1.6 nanometres thick. They also saw signs of a rare kind of magnetism.
🧒 In simple words: Scientists made a special material work even thinner than a strand of DNA. One day it could make memory chips use far less power.
🧩 Think of it like: Like a paint that used to stop working if you brushed it on too thin; now even a near-invisible coat still works.
How close to real use (0–100)16/100
Years to market: 6-10
Our judgment
🎢 Talk vs. real progress
Talk 40
Real progress 16
Some hype; watch for proof (our judgment)
⚖️ Patent / patents risk
Early science; little patent crowding.
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
In plain words: Scientists made a special material work even thinner than a strand of DNA. One day it could make memory chips use far less power.
📊 Stocks tied to this
Company
Stock code
How it is linked
Role
How sure we are
No listed company directly tied yet. Developer: Brown University and collaborators (Nature Materials). Watch for a listed partner or IPO💡.
Other possible winners: Memory and spintronics research labs (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Thin films of a material called bismuth ferrite usually stop working near the surface. That is called a "dead layer". The team removed it. The material still worked at about 1.6 nanometres thick. They also saw signs of a…
How close to real use (0–100) 16/100, stage 1 on our ladder.
score /100: half of how-ready score 16 + 0 (no big event expected soon) − 8×1 warning flag(s) · no listed stock code linked → ⚪ Too early: still early
⛔ Why not yet
Films are grown on a rare-earth DyScO3 crystal, not on standard silicon wafers
Altermagnetic order is shown by indirect signatures, which other groups still need to confirm
Timeline to impact: about 6-10 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Making it on normal silicon instead of a rare crystal, and a working memory cell.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More labs reproduce it.
Test memory cells in labs.
Still research unless it grows on silicon.
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Thin films of a material called bismuth ferrite usually stop working near the surface.
Who could win: Memory and spintronics research labs
🐻 What could go wrong
Films are grown on a rare-earth DyScO3 crystal, not on standard silicon wafers source
Altermagnetic order is shown by indirect signatures, which other groups still need to confirm source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
1/5
20%
4.0
Supply chain
1/5
10%
2.0
Corporate backing
1/5
10%
2.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
16/100
Talk score = 10×claim boldness (2/5) + 10×attention (2/5) = 40. Risk of a fight over who owns the idea: Low (crowding 2/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Making it on normal silicon instead of a rare crystal, and a working memory cell.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More labs reproduce it.
📋 Executive briefing teaser free
This kind of material could let voltage flip memory bits.
This kind of material could let voltage flip memory bits. That would use less energy. The paper shows it works at a size real devices need. It is basic science. Factories are far off.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Chips waste energy pushing current to flip bits. A material that flips magnetism with voltage could fix that, but it died when made thin. Brown just made it work at 1.6 nanometres. Source: Nature Materials.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Timeline steps are our judgment unless marked as company plan.
What they found: Altera showed its new Agilex 2X chips, led by Agilex 27. It claims 1.8 times more work per unit of power than before. They are for AI buildings, networks and small devices.
🧒 In simple words: Most chips do one job forever. This one can be re-programmed for new jobs. The new version uses less electricity.
🧩 Think of it like: A chip you can re-wire after you buy it, like Lego you can rebuild for each new AI job.
How close to real use (0–100)67/100
Years to market: 2-4
Our judgment
🎢 Talk vs. real progress
Talk 40
Real progress 67
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
FPGA patents are concentrated in AMD (Xilinx) and Altera. Some historic litigation.
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
In plain words: Most chips do one job forever. This one can be re-programmed for new jobs. The new version uses less electricity.
📊 Stocks tied to this
Company
Stock code
How it is linked
Role
How sure we are
No listed company directly tied yet. Developer: Altera (majority owned by Silver Lake; Intel minority stake). Watch for a listed partner or IPO💡.
Other possible winners: Altera, Network and telecom equipment makers (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Altera showed its new Agilex 2X chips, led by Agilex 27. It claims 1.8 times more work per unit of power than before. They are for AI buildings, networks and small devices.
How close to real use (0–100) 67/100, stage 2 on our ladder.
score /100: half of how-ready score 67 + 0 (no big event expected soon) · no listed stock code linked → ⚪ Too early: still early
⛔ Why not yet
Performance-per-watt figure is vendor-claimed, not independently tested
For pure AI maths, FPGAs usually lose to GPUs and custom chips on cost per task
Timeline to impact: about 2-4 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Buyers choosing Agilex 27, and its shipping dates.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Early customer designs.
In network gear and AI servers.
Replaced by the next generation.
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Altera showed its new Agilex 2X chips, led by Agilex 27.
Who could win: Altera, Network and telecom equipment makers
🐻 What could go wrong
Performance-per-watt figure is vendor-claimed, not independently tested source
For pure AI maths, FPGAs usually lose to GPUs and custom chips on cost per task (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
3/5 — Customer samples / qualification
55%
24.8
Manufacturability
5/5
20%
20.0
Supply chain
5/5
10%
10.0
Corporate backing
4/5
10%
8.0
Independent groups
4/5
5%
4.0
How close to real use (0–100)
67/100
Talk score = 10×claim boldness (2/5) + 10×attention (2/5) = 40. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Buyers choosing Agilex 27, and its shipping dates.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Early customer designs.
📋 Executive briefing teaser free
Re-programmable chips fill gaps where needs change too fast for custom chips.
Re-programmable chips fill gaps where needs change too fast for custom chips. The number comes from the maker. Altera is private, so you cannot buy its shares on a stock market.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Some chips can be re-wired after they ship. Altera's new Agilex 2X claims 1.8× more work per watt for AI networks. Source: Altera.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Timeline steps are our judgment unless marked as company plan.
What they found: Ceva showed new NeuPro chip plans for AI. It claims twice the work in the same space. It also claims 50% better power use. Ceva is moving from small devices into big AI computers. Early access is open now. Full release is planned for April–June 2027.
🧒 In simple words: Ceva draws plans for the "thinking" part of a chip. Other companies use the plans. The new plans do twice the work in the same space.
🧩 Think of it like: Ceva sells blueprints, not buildings. Chip makers license the AI-engine design and build it into their own chips.
How close to real use (0–100)66/100
Years to market: 2-4
Our judgment
🎢 Talk vs. real progress
Talk 50
Real progress 66
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
NPU patents is crowded (Arm, Cadence, Synopsys, many startups).
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
Other possible winners: Ceva, Chip makers licensing NPU patents (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Ceva showed new NeuPro chip plans for AI. It claims twice the work in the same space. It also claims 50% better power use. Ceva is moving from small devices into big AI computers. Early access is open now. Full release i…
How close to real use (0–100) 66/100, stage 2 on our ladder.
score /100: half of how-ready score 66 + 0 (no big event expected soon)
⛔ Why not yet
General availability not until Q2 2027; figures are vendor-claimed
Crowded market against Arm, Cadence, Synopsys and in-house designs
Timeline to impact: about 2-4 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Named companies using the new plans, and the April–June 2027 full release.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Early-access licensees.
First chips using it ship.
Royalties if widely adopted.
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Ceva showed new NeuPro chip plans for AI.
Who could win: Ceva, Chip makers licensing NPU patents
🐻 What could go wrong
General availability not until Q2 2027; figures are vendor-claimed source
Crowded market against Arm, Cadence, Synopsys and in-house designs (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
3/5 — Customer samples / qualification
55%
24.8
Manufacturability
5/5
20%
20.0
Supply chain
5/5
10%
10.0
Corporate backing
3/5
10%
6.0
Independent groups
5/5
5%
5.0
How close to real use (0–100)
66/100
Talk score = 10×claim boldness (3/5) + 10×attention (2/5) = 50. Risk of a fight over who owns the idea: Medium (crowding 4/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Named companies using the new plans, and the April–June 2027 full release.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Early-access licensees.
📋 Executive briefing teaser free
Ready-made plans let many companies add AI without starting from zero.
Ready-made plans let many companies add AI without starting from zero. The gains are Ceva's own claims. Ceva gets paid small fees later, as chips sell, often years later.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Not every company can design an AI chip. Ceva sells the design. Its new NeuPro claims double the work per area. Ships to licensees in 2027. Source: Ceva.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
🔬 A startup is building chips that do AI maths with light.
🧩 Like sorting mail by shining a torch through a stencil instead of reading every envelope.
🚀 Why it matters
📈 How close to real use (0–100) 31/100 · Prototype
⏳ 3-6 years to market (our judgment)
🎯 This light-based AI chip makes no money yet.
💰 Who gains
Tower SemiconductorUS stock market: TSEM
🏷 Neurophos (private, US)
🏷 GlobalFoundries GFS (US, photonics foundry)
🏷 Tower Semiconductor TSEM (Israel)
⚠️ Risks
❌ Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores.
Step 2Pilot units in a few data centres, if tests go well
Step 3Could be a niche accelerator for specific AI jobs
Timeline steps are our judgment unless marked as company plan.
What they found: A startup is building chips that do AI maths with light. The light passes through tiny patterns called metamaterials. It is working with MIT on the software to run them.
🧒 In simple words: One day, some computers may think with light instead of electricity. That keeps them cool and fast.
🧩 Think of it like: Like sorting mail by shining a torch through a stencil instead of reading every envelope.
How close to real use (0–100)31/100
Years to market: 3-6
Our judgment
🎢 Talk vs. real progress
Talk 80
Real progress 31
Mostly hype so far (our judgment)
⚖️ Patent / patents risk
Neurophos holds granted US and international patents on its optical design. Other optical-compute firms (e.g. Lightmatter) hold their own patents, so overlap disputes are possible. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
Other possible winners: Neurophos (private, US), GlobalFoundries GFS (US, photonics foundry), Tower Semiconductor TSEM (Israel) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
A startup is building chips that do AI maths with light. The light passes through tiny patterns called metamaterials. It is working with MIT on the software to run them.
How close to real use (0–100) 31/100, stage 2 on our ladder.
score /100: half of how-ready score 31 + 0 (no big event expected soon) − 8×1 warning flag(s)
⛔ Why not yet
Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores.
SPIE paper: data converters can be over 80% of the conversion power budget in photonic AI accelerators.
Timeline to impact: about 3-6 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: First chip samples sent to buyers, or news of a chip-factory partner.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Lab demos and first test chips
Pilot units in a few data centres, if tests go well
Could be a niche accelerator for specific AI jobs
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
A startup is building chips that do AI maths with light.
Who could win: Neurophos (private, US), GlobalFoundries GFS (US, photonics foundry), Tower Semiconductor TSEM (Israel)
🐻 What could go wrong
Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores. source
SPIE paper: data converters can be over 80% of the conversion power budget in photonic AI accelerators. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
2/5 — Working prototype chip
55%
13.8
Manufacturability
2/5
20%
8.0
Supply chain
1/5
10%
2.0
Corporate backing
2/5
10%
4.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
31/100
Talk score = 10×claim boldness (4/5) + 10×attention (4/5) = 80. Risk of a fight over who owns the idea: Medium (crowding 4/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: First chip samples sent to buyers, or news of a chip-factory partner.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Lab demos and first test chips
📋 Executive briefing teaser free
This light-based AI chip makes no money yet.
This light-based AI chip makes no money yet. It has patents and an MIT software partner. Look for news that the first chip design is finished. Chip factories with light-chip lines may gain either way.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
What if your AI chip ran on light? A US startup called Neurophos is doing AI math by shining light through microscopic patterns. Less heat, more speed. It's early: no product yet. But they hold patents and just teamed up with MIT. If it works, photonics foundries win first. Source links on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Timeline steps are our judgment unless marked as company plan.
What they found: Volantis showed a plan to put light links in the base layer under an AI chip. That layer also sits under the chip's memory. Data would then move by light.
🧒 In simple words: Chips wait for memory, like people waiting in a long queue. This idea builds a fast light path to bring data quicker.
🧩 Think of it like: Replacing a crowded two-lane road between factory and warehouse with a bullet train.
How close to real use (0–100)32/100
Years to market: 4-8
Our judgment
🎢 Talk vs. real progress
Talk 80
Real progress 32
Mostly hype so far (our judgment)
⚖️ Patent / patents risk
Volantis has filed patents (see assignee profile). Big players like Nvidia, Broadcom and Ayar Labs hold large optical I/O portfolios, so licensing may be needed. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
Volantis (private)Coherent COHR (US)Lumentum LITE (US)STMicro STM (EU)
Other possible winners: Volantis (private), Coherent COHR (US), Lumentum LITE (US), STMicro STM (EU) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Volantis showed a plan to put light links in the base layer under an AI chip. That layer also sits under the chip's memory. Data would then move by light.
How close to real use (0–100) 32/100, stage 1 on our ladder.
score /100: half of how-ready score 32 + 0 (no big event expected soon) − 8×2 warning flag(s)
⛔ Why not yet
Nvidia networking SVP: 'If you can use copper, use copper' — optics only where distance forces it.
SemiAnalysis: Nvidia is going to 'extreme lengths' to stay on copper for scale-up through 2027-2028; limited first-wave CPO adoption expected.
Timeline to impact: about 4-8 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Test results from a working sample, and named buyers trying it.
A named paying customer or order
Independent replication / outside test results
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Prototype shown to partners
Early customer trials
Possibly inside some AI servers
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Volantis showed a plan to put light links in the base layer under an AI chip.
Who could win: Volantis (private), Coherent COHR (US), Lumentum LITE (US)
🐻 What could go wrong
Nvidia networking SVP: 'If you can use copper, use copper' — optics only where distance forces it. source
SemiAnalysis: Nvidia is going to 'extreme lengths' to stay on copper for scale-up through 2027-2028; limited first-wave CPO adoption expected. source
Plan revealed, no shipping product disclosed yet (our reading of the source). (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
3/5
20%
12.0
Supply chain
3/5
10%
6.0
Corporate backing
2/5
10%
4.0
Independent groups
4/5
5%
4.0
How close to real use (0–100)
32/100
Talk score = 10×claim boldness (4/5) + 10×attention (4/5) = 80. Risk of a fight over who owns the idea: Medium (crowding 4/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Test results from a working sample, and named buyers trying it.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Prototype shown to partners
📋 Executive briefing teaser free
It is an early idea to fix slow memory links.
It is an early idea to fix slow memory links. The room to gain is big, but so is the risk. Laser and light-part makers gain whichever start-up wins.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
AI chips are starving, waiting for data from memory. Volantis wants to fix that with light: an optical highway built right under the chip. It's just a plan for now. But if it works, laser makers like Coherent and Lumentum get a new market. Links on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
🔬 Researchers packed several pieces of data into one memory cell.
🧩 Like one shelf that holds several books layered as transparent sheets, still readable.
🚀 Why it matters
📈 How close to real use (0–100) 19/100 · Paper
⏳ 6-10 years to market (our judgment)
🎯 IBM is filing patents along with this research.
💰 Who gains
MicronUS stock market: MU
🏷 IBM (US)
🏷 Micron MU (US)
🏷 Weebit Nano WBT (Australia, small)
⚠️ Risks
❌ System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted.
Timeline steps are our judgment unless marked as company plan.
What they found: Researchers packed several pieces of data into one memory cell. They could still read each piece back. This cuts the memory AI search needs.
🧒 In simple words: Imagine one drawer that holds three things at once. You can still pick each one out.
🧩 Think of it like: Like one shelf that holds several books layered as transparent sheets, still readable.
How close to real use (0–100)19/100
Years to market: 6-10
Our judgment
🎢 Talk vs. real progress
Talk 40
Real progress 19
Some hype; watch for proof (our judgment)
⚖️ Patent / patents risk
IBM has both the paper and a matching patent application plus older granted patents: strong patents wall. Anyone copying the method may need a licence. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
IBM (US)Micron MU (US)Weebit Nano WBT (Australia, small)
Other possible winners: IBM (US), Micron MU (US), Weebit Nano WBT (Australia, small) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Researchers packed several pieces of data into one memory cell. They could still read each piece back. This cuts the memory AI search needs.
How close to real use (0–100) 19/100, stage 1 on our ladder.
score /100: half of how-ready score 19 + 0 (no big event expected soon) − 8×1 warning flag(s)
⛔ Why not yet
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted.
Timeline to impact: about 6-10 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: A test on IBM's real AI chip hardware.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More papers and patents
Tested on IBM's research chips
Maybe in specialised memory products
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Researchers packed several pieces of data into one memory cell.
Who could win: IBM (US), Micron MU (US), Weebit Nano WBT (Australia, small)
🐻 What could go wrong
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
0/5 — Paper / simulation only
55%
0.0
Manufacturability
2/5
20%
8.0
Supply chain
2/5
10%
4.0
Corporate backing
3/5
10%
6.0
Independent groups
1/5
5%
1.0
How close to real use (0–100)
19/100
Talk score = 10×claim boldness (2/5) + 10×attention (2/5) = 40. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: A test on IBM's real AI chip hardware.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More papers and patents
📋 Executive briefing teaser free
IBM is filing patents along with this research.
IBM is filing patents along with this research. That hints it wants to sell it one day. It is years away. Memory makers may care most.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Imagine one memory cell holding several numbers at once. IBM just published that trick, and filed a patent on it the same season. That's a sign they plan to use it. It's years away, but memory makers should pay attention. Sources on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
🏷 Kaynes KAYNES (India (NSE)) (India, OSAT partner ecosystem)
⚠️ Risks
❌ System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted.
Timeline steps are our judgment unless marked as company plan.
What they found: Micron split each AI number across several memory cells. Averaging them cancels out errors. Those errors make this type of chip unreliable.
🧒 In simple words: If one friend guesses wrong, ask three and take the middle answer. Micron does that inside a chip.
🧩 Think of it like: Asking three friends to estimate a distance and averaging: fewer big mistakes.
How close to real use (0–100)35/100
Years to market: 4-8
Our judgment
🎢 Talk vs. real progress
Talk 30
Real progress 35
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
Splitting/redundancy ideas are widely published; patent risk looks moderate. Micron may file follow-ups. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
Micron MU (US, India plant in Gujarat)Kaynes KAYNES (India (NSE)) (India, OSAT partner ecosystem)
Other possible winners: Micron MU (US, India plant in Gujarat), Kaynes KAYNES (India (NSE)) (India, OSAT partner ecosystem) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Micron split each AI number across several memory cells. Averaging them cancels out errors. Those errors make this type of chip unreliable.
How close to real use (0–100) 35/100, stage 1 on our ladder.
score /100: half of how-ready score 35 + 0 (no big event expected soon) − 8×1 warning flag(s)
⛔ Why not yet
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted.
Timeline to impact: about 4-8 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Whether Micron shows it on a real chip or in its product plans.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Follow-up papers
Shown on real Micron test silicon
Could ship inside analog-AI memory products
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Micron split each AI number across several memory cells.
Who could win: Micron MU (US, India plant in Gujarat), Kaynes KAYNES (India (NSE)) (India, OSAT partner ecosystem)
🐻 What could go wrong
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
0/5 — Paper / simulation only
55%
0.0
Manufacturability
4/5
20%
16.0
Supply chain
4/5
10%
8.0
Corporate backing
4/5
10%
8.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
35/100
Talk score = 10×claim boldness (1/5) + 10×attention (2/5) = 30. Risk of a fight over who owns the idea: Low (crowding 2/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Whether Micron shows it on a real chip or in its product plans.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Follow-up papers
📋 Executive briefing teaser free
A big memory maker is fixing errors in this chip type.
A big memory maker is fixing errors in this chip type. That hints it is moving toward real products. Micron gains most directly.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Analog AI chips are super efficient but sloppy. Micron found a simple fix: spread each number across several cells, like averaging friends' guesses. When a memory giant solves accuracy, products tend to follow. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More lab records
Included in next-gen 3.2T optics plans
Common in data-centre cables
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
A tiny silicon part switches light on and off very fast.
Who could win: Marvell MRVL (US), Broadcom AVGO (US), Coherent COHR (US)
🐻 What could go wrong
Nvidia networking SVP: 'If you can use copper, use copper' — optics only where distance forces it. source
SemiAnalysis: Nvidia is going to 'extreme lengths' to stay on copper for scale-up through 2027-2028; limited first-wave CPO adoption expected. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
4/5
20%
16.0
Supply chain
4/5
10%
8.0
Corporate backing
1/5
10%
2.0
Independent groups
4/5
5%
4.0
How close to real use (0–100)
36/100
Talk score = 10×claim boldness (3/5) + 10×attention (1/5) = 40. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Use in the next faster light-link products (1.6T and 3.2T).
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More lab records
📋 Executive briefing teaser free
Speed per light lane matters a lot in data centres.
Speed per light lane matters a lot in data centres. This is a lab result today. But it fits the plans of companies that make light links.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Data centres talk with light. A new chip-sized switch flips light 200 billion times a second on one colour. Faster switches mean cheaper, cooler AI clusters. Optics makers like Marvell and Coherent are the ones to watch. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Step 3Still mostly research; mass use likely in the 2030s
Timeline steps are our judgment unless marked as company plan.
What they found: imec tested stacked switches made of atom-thin materials. Do they beat silicon in chips from about 2026? Not yet. Losses in the wiring eat up the gains.
🧒 In simple words: This idea makes chip switches from sheets one atom thick. Then it stacks them like pancakes.
🧩 Think of it like: A thinner wall doesn't help if the doors and hallways are the bottleneck.
How close to real use (0–100)16/100
Years to market: 6-10
Our judgment
🎢 Talk vs. real progress
Talk 50
Real progress 16
Some hype; watch for proof (our judgment)
⚖️ Patent / patents risk
CFET structures are heavily patented by TSMC, Intel, Samsung and IBM. Any newcomer faces dense patents. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
ASML (EU)Applied Materials AMAT (US)Lam Research LRCX (US)Aixtron AIXA (EU, 2D-material tools)
Other possible winners: ASML (EU), Applied Materials AMAT (US), Lam Research LRCX (US), Aixtron AIXA (EU, 2D-material tools) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
imec tested stacked switches made of atom-thin materials. Do they beat silicon in chips from about 2026? Not yet. Losses in the wiring eat up the gains.
How close to real use (0–100) 16/100, stage 1 on our ladder.
score /100: half of how-ready score 16 + 0 (no big event expected soon) − 8×1 warning flag(s)
⛔ Why not yet
No fab today grows 2D-material layers at wafer scale with chip-grade yield (our judgment; no single source).
Timeline to impact: about 6-10 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: imec or chip-factory papers at the IEDM meeting in December on fixing the wiring losses.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More lab tests on contacts
Research pilot lines
Still mostly research; mass use likely in the 2030s
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
imec tested stacked switches made of atom-thin materials.
Who could win: ASML (EU), Applied Materials AMAT (US), Lam Research LRCX (US)
🐻 What could go wrong
No fab today grows 2D-material layers at wafer scale with chip-grade yield (our judgment; no single source). (judgment)
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
0/5
20%
0.0
Supply chain
1/5
10%
2.0
Corporate backing
3/5
10%
6.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
16/100
Talk score = 10×claim boldness (3/5) + 10×attention (2/5) = 50. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: imec or chip-factory papers at the IEDM meeting in December on fixing the wiring losses.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More lab tests on contacts
📋 Executive briefing teaser free
This paper is a reality check.
This paper is a reality check. These chips are a 2030s story. Chip-tool makers gain either way. Every new chip design needs new tools.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Chips are running out of room, so engineers want to stack transistors made of material just one atom thick. Europe's top chip lab, imec, tested it, and says: not yet. That honesty is useful. The safe winners? The toolmakers. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Timeline steps are our judgment unless marked as company plan.
What they found: Tool makers are teaming up to sell machines that fuse chips copper to copper. But SK hynix says it won't be ready for its next memory type (HBM4E).
🧒 In simple words: Chips are usually joined with tiny solder balls. Here they are pressed so the copper fuses, like welding.
🧩 Think of it like: Welding Lego bricks instead of snapping them: tighter, but harder to do right.
How close to real use (0–100)80/100
Years to market: 0-2
Our judgment
🎢 Talk vs. real progress
Talk 50
Real progress 80
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
Besi/Applied, Adeia (licensor, ADEA) and TSMC hold key hybrid-bonding patents. Adeia actively licenses: a royalty stream, and a lawsuit risk for others. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More orders for bonding machines
Used in some memory and logic stacks
Standard for top AI chips
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Tool makers are teaming up to sell machines that fuse chips copper to copper.
Who could win: Besi BESI (Netherlands) (EU), Applied Materials AMAT (US), Adeia ADEA (US, small, patents licensor)
🐻 What could go wrong
SK hynix says hybrid bonding won't be ready for HBM4E memory stacks. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
4/5 — Volume production or orders
55%
41.2
Manufacturability
4/5
20%
16.0
Supply chain
4/5
10%
8.0
Corporate backing
5/5
10%
10.0
Independent groups
5/5
5%
5.0
How close to real use (0–100)
80/100
Talk score = 10×claim boldness (2/5) + 10×attention (3/5) = 50. Risk of a fight over who owns the idea: High (crowding 4/5, litigation 4/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Which memory type uses it first (maybe HBM5), and new orders at Besi.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More orders for bonding machines
📋 Executive briefing teaser free
This joining method now has real orders.
This joining method now has real orders. But it slipped for the next memory type. Older tool makers gain for now. Besi and AMAT may gain later. Adeia holds patents here.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Next-gen AI chips are glued together atom-close using copper. Toolmakers Applied Materials and Besi are teaming up to sell the machines. But memory giant SK hynix says it's not ready yet. Delay now, big wave later. Links on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
13. TSMC and others swap plastic for glass under AI chips
In plain words: The board under a chip may switch from plastic to glass. Glass stays flat and doesn't bend.
📰What happenedThe board under a chip may switch from plastic to glass. Glass stays flat and doesn't bend.
💡Why you should careGlass bases are moving from the lab to small test factories. Base-board and glass makers are ready.
🔭What happens nextWatch for: How many good parts test factories make, and the first AI product with glass. Oct 29 update: Lam Research says glass helpers are used today. But glass as the main base is still early. So no stage change.
🧩 Building a skyscraper on granite instead of plywood.
Other possible winners: Ibiden 4062 (Japan), Samsung Electro-Mechanics 009150 (Korea), SKC 011790 (Korea), Corning GLW (US), 3D Glass Solutions (India plant, Odisha) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Several firms are testing glass as the base under giant AI chips. Glass stays flat. It also lets wires sit closer together.
How close to real use (0–100) 52/100, stage 2 on our ladder.
score /100: half of how-ready score 52 + 0 (no big event expected soon)
⛔ Why not yet
Tom's Hardware: glass-core substrates are 'still not in a single commercial product' and every timeline has slipped.
TrendForce/The Elec: Samsung Electro-Mechanics order schedule pushed back at least 3 times; mass production possibly 2028+ (company disputes a failed qualification).
Timeline to impact: about 2-4 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: How many good parts test factories make, and the first AI product with glass. Oct 29 update: Lam Research says glass helpers are used today. But glass as the main base is still early. So no stage change.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Pilot lines and samples
First AI products on glass, if yields are good
Common in big AI chips
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Several firms are testing glass as the base under giant AI chips.
Who could win: Ibiden 4062 (Japan), Samsung Electro-Mechanics 009150 (Korea), SKC 011790 (Korea)
🐻 What could go wrong
Tom's Hardware: glass-core substrates are 'still not in a single commercial product' and every timeline has slipped. source
TrendForce/The Elec: Samsung Electro-Mechanics order schedule pushed back at least 3 times; mass production possibly 2028+ (company disputes a failed qualification). source
Lam Research (Oct 29): permanent glass interposers and glass-core substrates are still at earlier stages; manufacturability and standards must be solved first source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
3/5 — Customer samples / qualification
55%
24.8
Manufacturability
2/5
20%
8.0
Supply chain
2/5
10%
4.0
Corporate backing
5/5
10%
10.0
Independent groups
5/5
5%
5.0
How close to real use (0–100)
52/100
Talk score = 10×claim boldness (3/5) + 10×attention (4/5) = 70. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: How many good parts test factories make, and the first AI product with glass. Oct 29 update: Lam Research says glass helpers are used today. But glass as the main base is still early. So no stage change.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Pilot lines and samples
📋 Executive briefing teaser free
Glass bases are moving from the lab to small test factories.
Glass bases are moving from the lab to small test factories. Base-board and glass makers are ready. India joins through 3DGS's project in Odisha.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
The board under an AI chip may soon be made of glass. It stays flat and fits more wires. TSMC is testing it with Japan's Ibiden; Samsung and BOE are racing too. Even India has a glass-packaging plant coming. Links on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
❌ BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny.
Timeline steps are our judgment unless marked as company plan.
What they found: A brain-style chip design can train itself. It was shown on a chip that can be re-programmed (FPGA). It beat Nvidia's Jetson Orin by 1.9 to 3.3 times. It also moved far less memory data than an A100 while training.
🧒 In simple words: This chip works like a tiny brain. It can learn new things by itself, without the internet.
🧩 Think of it like: A student who learns while working instead of going back to school.
How close to real use (0–100)28/100
Years to market: 4-8
Our judgment
🎢 Talk vs. real progress
Talk 60
Real progress 28
Some hype; watch for proof (our judgment)
⚖️ Patent / patents risk
Neuromorphic patents are held by Intel (Loihi), IBM and BrainChip; academic designs can face overlap. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
Other possible winners: BrainChip BRN (Australia) (Australia, small), Lattice LSCC (US, FPGAs), AMD (Xilinx FPGAs) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
A brain-style chip design can train itself. It was shown on a chip that can be re-programmed (FPGA). It beat Nvidia's Jetson Orin by 1.9 to 3.3 times. It also moved far less memory data than an A100 while training.
How close to real use (0–100) 28/100, stage 1 on our ladder.
score /100: half of how-ready score 28 + 0 (no big event expected soon) − 8×1 warning flag(s)
⛔ Why not yet
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny.
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted.
Timeline to impact: about 4-8 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: A real chip made in a factory, and device makers signing up.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
More FPGA demos
First real chips (tape-out)
Possible use in robots and cameras
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
A brain-style chip design can train itself.
Who could win: BrainChip BRN (Australia) (Australia, small), Lattice LSCC (US, FPGAs), AMD (Xilinx FPGAs)
🐻 What could go wrong
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny. source
System-level review: specialised analog/physical accelerators can show 'little or no advantage' once conversion, control and memory access are counted. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
3/5
20%
12.0
Supply chain
2/5
10%
4.0
Corporate backing
1/5
10%
2.0
Independent groups
4/5
5%
4.0
How close to real use (0–100)
28/100
Talk score = 10×claim boldness (3/5) + 10×attention (3/5) = 60. Risk of a fight over who owns the idea: Low (crowding 2/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: A real chip made in a factory, and device makers signing up.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: More FPGA demos
📋 Executive briefing teaser free
Devices learning on their own is a real gap.
Devices learning on their own is a real gap. Today it is only a demo. Small makers of brain-style and re-programmable chips are linked.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Most AI learns in giant data centres. This brain-like chip learns on the spot, and it beat Nvidia's Jetson by up to 3 times in tests. It's an FPGA demo, not a product. Keep watching small players like BrainChip. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
🧩 A Swiss-army-knife worker who becomes a gatekeeper, clerk or manager as needed.
🚀 Why it matters
📈 How close to real use (0–100) 16/100 · Paper
⏳ 6-10 years to market (our judgment)
🎯 This is pure research.
💰 Who gains
No listed company that mostly does this linked (private or small part of a big company)
🏷 Ferroelectric Memory Co (private, Germany)
🏷 GlobalFoundries GFS (US)
⚠️ Risks
❌ BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny.
Timeline steps are our judgment unless marked as company plan.
What they found: An atom-thin switch can change shape and role. It can act as logic or like parts of a brain cell. It throws out useless data before wasting energy on it.
🧒 In simple words: One tiny switch can change jobs. Sometimes it is a calculator. Sometimes it acts like a brain cell.
🧩 Think of it like: A Swiss-army-knife worker who becomes a gatekeeper, clerk or manager as needed.
In plain words: One tiny switch can change jobs. Sometimes it is a calculator. Sometimes it acts like a brain cell.
📊 Stocks tied to this
Company
Stock code
How it is linked
Role
How sure we are
No listed company directly tied yet. Developer: National University of Singapore-led team. Watch for a listed partner or IPO💡.
Other possible winners: Ferroelectric Memory Co (private, Germany), GlobalFoundries GFS (US) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
An atom-thin switch can change shape and role. It can act as logic or like parts of a brain cell. It throws out useless data before wasting energy on it.
How close to real use (0–100) 16/100, stage 1 on our ladder.
score /100: half of how-ready score 16 + 0 (no big event expected soon) − 8×1 warning flag(s) · no listed stock code linked → ⚪ Too early: still early
⛔ Why not yet
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny.
Timeline to impact: about 6-10 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Results from making it on full-size chip wafers.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Bigger arrays in labs
Wafer-scale tests
Still research
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
An atom-thin switch can change shape and role.
Who could win: Ferroelectric Memory Co (private, Germany), GlobalFoundries GFS (US)
🐻 What could go wrong
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
1/5
20%
4.0
Supply chain
1/5
10%
2.0
Corporate backing
1/5
10%
2.0
Independent groups
2/5
5%
2.0
How close to real use (0–100)
16/100
Talk score = 10×claim boldness (2/5) + 10×attention (1/5) = 30. Risk of a fight over who owns the idea: Low (crowding 2/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Results from making it on full-size chip wafers.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Bigger arrays in labs
📋 Executive briefing teaser free
This is pure research.
This is pure research. The direction looks good. No listed company mostly does this yet.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
What if one tiny transistor could switch jobs: logic one moment, brain cell the next? Singapore researchers built exactly that from a material one atom thick. Decade-away science, but it's how the future starts. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
🔬 Light-based parts act like brain cells and their links.
🧩 Swapping a postal service for a laser pointer between two windows.
🚀 Why it matters
📈 How close to real use (0–100) 16/100 · Paper
⏳ 6-10 years to market (our judgment)
🎯 The lab gains look big.
💰 Who gains
No listed company that mostly does this linked (private or small part of a big company)
🏷 Chinese photonics foundries (not easily investable)
🏷 Tower TSEM (Israel)
⚠️ Risks
❌ Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores.
Step 3Possible edge-vision products, mostly in China
Timeline steps are our judgment unless marked as company plan.
What they found: Light-based parts act like brain cells and their links. In their tests on picture tasks, delay fell 1,176 times. Energy fell 239 times.
🧒 In simple words: A brain-style chip uses light. In tests it was over a thousand times faster at seeing pictures.
🧩 Think of it like: Swapping a postal service for a laser pointer between two windows.
How close to real use (0–100)16/100
Years to market: 6-10
Our judgment
🎢 Talk vs. real progress
Talk 70
Real progress 16
Mostly hype so far (our judgment)
⚖️ Patent / patents risk
Chinese groups patent heavily; export-control limits could restrict Western use. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
Chinese photonics foundries (not easily investable)Tower TSEM (Israel)
In plain words: A brain-style chip uses light. In tests it was over a thousand times faster at seeing pictures.
📊 Stocks tied to this
Company
Stock code
How it is linked
Role
How sure we are
No listed company directly tied yet. Developer: Fudan University-led team (China). Watch for a listed partner or IPO💡.
Other possible winners: Chinese photonics foundries (not easily investable), Tower TSEM (Israel) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
Light-based parts act like brain cells and their links. In their tests on picture tasks, delay fell 1,176 times. Energy fell 239 times.
How close to real use (0–100) 16/100, stage 1 on our ladder.
U=-8: half of how-ready score 16 + 0 (no big event expected soon) − 8×2 warning flag(s) · no listed stock code linked → ⚪ Too early: still early
⛔ Why not yet
Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores.
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny.
Timeline to impact: about 6-10 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Other teams repeating it, and bigger versions.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Add the stock codes to a list to follow and check the triggers below each week.
Move to Build only if 2+ upgrade triggers happen.
🗓 Timeline
Other labs try to repeat it
Bigger prototype chips
Possible edge-vision products, mostly in China
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
Light-based parts act like brain cells and their links.
Who could win: Chinese photonics foundries (not easily investable), Tower TSEM (Israel)
🐻 What could go wrong
Nature Photonics review (2026): in the cloud, losses and light-to-electric conversions 'dominate power' and have so far prevented commercial photonic tensor cores. source
BrainChip, a listed neuromorphic pioneer, reported only US$1.22M revenue and a US$12.0M loss for H1 2026 — commercial demand for brain-like chips is still tiny. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
1/5 — Lab demo
55%
5.5
Manufacturability
1/5
20%
4.0
Supply chain
1/5
10%
2.0
Corporate backing
1/5
10%
2.0
Independent groups
3/5
5%
3.0
How close to real use (0–100)
16/100
Talk score = 10×claim boldness (5/5) + 10×attention (2/5) = 70. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 1/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Other teams repeating it, and bigger versions.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Other labs try to repeat it
📋 Executive briefing teaser free
The lab gains look big.
The lab gains look big. Politics may keep this tech inside China. Keep watching whether others can repeat it.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
A Chinese lab says its light-based brain chip is over a thousand times faster on vision tasks. Big claim, lab test. If it holds up, photonics becomes the brain of edge AI. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Step 2Over 4x capacity planned by 2027 (company plan)
Step 3Next version (TSV) for co-packaged optics
Timeline steps are our judgment unless marked as company plan.
What they found: ST says its light-chip process (PIC100) is now in high-volume production. It runs on 300 mm lines for giant cloud companies. It makes 800G and 1.6T light links. ST plans to more than quadruple capacity by 2027.
🧒 In simple words: Chips that talk using light are already made in big factories right now.
🧩 Think of it like: Like the printing press replacing hand-copied books: same words, but suddenly thousands of copies, cheaply.
How close to real use (0–100)84/100
Years to market: 0-2
Our judgment
🎢 Talk vs. real progress
Talk 30
Real progress 84
Real progress is catching up with the talk (our judgment)
⚖️ Patent / patents risk
ST licenses/owns its process; competitors (Intel, TSMC COUPE, GlobalFoundries) have their own. Low infringement risk for buyers. (Our judgment)
🔒 Ownership-fight noteIncluded in Pro & InstitutionalUnlock Premium
🏆 Who could win
STMicro STM (EU)Coherent COHR (US)Lumentum LITE (US)
Other possible winners: STMicro STM (EU), Coherent COHR (US), Lumentum LITE (US) (our judgment)
⚠️ Who could lose
Makers of older copper/electrical interconnect and packaging may lose share (our judgment)
✅ Why now
ST says its light-chip process (PIC100) is now in high-volume production. It runs on 300 mm lines for giant cloud companies. It makes 800G and 1.6T light links. ST plans to more than quadruple capacity by 2027.
How close to real use (0–100) 84/100, stage 3 on our ladder.
score /100: half of how-ready score 84 + 0 (no big event expected soon)
⛔ Why not yet
Nvidia networking SVP: 'If you can use copper, use copper' — optics only where distance forces it.
SemiAnalysis: Nvidia is going to 'extreme lengths' to stay on copper for scale-up through 2027-2028; limited first-wave CPO adoption expected.
Timeline to impact: about 0-2 years (our judgment)
⬆️ Upgrade triggers
Concrete date or release: Named cloud-company buyers, and ST sharing sales numbers.
A named paying customer or order
Independent replication / outside test results💡
How close to real use (0–100) reaching 70+ with a dated big upcoming event within 90 days (our 🔥 rule)
⬇️ Downgrade signs
Delay or missed milestone (our judgment)
Independent tests fail to confirm the claim (our judgment)
A rival leapfrogs or rules/regulators block it (our judgment)
Only a small, staged position (e.g. split buys over several months).
Size so a big drop wouldn't hurt you.
Stop adding if a downgrade sign appears.
🗓 Timeline
Capacity grows; more customers
Over 4x capacity planned by 2027 (company plan)
Next version (TSV) for co-packaged optics
📖 What these words mean
How close to real use (0–100): our 0–100 score for how close an idea is to real products and sales. Higher = closer.
How sure we are: how confident we are in a company's link to the idea (high, medium, low).
Stock code: the short code used to find a company's share on an exchange, e.g. NVDA.
Big upcoming event / next big event: a dated event (launch, approval, results) that could move the story.
how soon: our badge for how time-sensitive it is: 🟢 moving now sign, ⏳ promising, ⚪ Too early: still early.
🇮🇳 Indian investor angle
NSE💡: KAYNES and CGPOWER (both started chip assembly/test, OSAT, production at Sanand in 2026); US names via Nasdaq-100 or US-tech funds under the LRS route (RBI scheme for investing abroad). (our judgment)
How it is linked notes are qualitative; we don't quote % of revenue unless sourced.
🐂 Good case
ST says its light-chip process (PIC100) is now in high-volume production.
Who could win: STMicro STM (EU), Coherent COHR (US), Lumentum LITE (US)
🐻 What could go wrong
Nvidia networking SVP: 'If you can use copper, use copper' — optics only where distance forces it. source
SemiAnalysis: Nvidia is going to 'extreme lengths' to stay on copper for scale-up through 2027-2028; limited first-wave CPO adoption expected. source
📐 How we scored this
Criterion
Score
Weight
Points
Evidence level
4/5 — Volume production or orders
55%
41.2
Manufacturability
5/5
20%
20.0
Supply chain
4/5
10%
8.0
Corporate backing
5/5
10%
10.0
Independent groups
5/5
5%
5.0
How close to real use (0–100)
84/100
Talk score = 10×claim boldness (1/5) + 10×attention (2/5) = 30. Risk of a fight over who owns the idea: Medium (crowding 3/5, litigation 2/5). Sub-scores are our judgment from the linked sources. Full method
👀 What to watch next: Named cloud-company buyers, and ST sharing sales numbers.
⏰ Time Machinedrag: 1 → 3 → 5 years (our scenario, not a forecast)
In 1 year: Capacity grows; more customers
📋 Executive briefing teaser free
This shows light-chip research can turn into sales.
This shows light-chip research can turn into sales. ST is already making lots. This is the near-term part of the light-chip story.
🔒 Full executive briefingIncluded in Pro & InstitutionalUnlock Premium
🎬 30-second video script Premium
Remember those light-based chip papers? Here's one that made it: ST now mass-produces silicon photonics for data centres. Paper to profit, in action. Link on our page.
🔒 Video script with citationsIncluded in Pro & InstitutionalUnlock Premium
Step 3 of 3 · More: Did you know, sign-up and word list. (tap to open)
💡 Did you know? Patents are often filed months before the paper appears. When a company files a patent and a paper on the same idea (like IBM here), it usually means they plan to sell it.
📬 Get tomorrow's breakthroughs first
New papers and patents, decoded in plain English, every morning.
