Lumivest
Sign in
HomeETFsWatchNewsHistorySettings
Sign in

News feed

AI-infrastructure signals from the last 24 hours, ranked by significance.

19 of 19 · 24h window
FilterAllBearishBullishSubstackAuto-refreshes each cycle

Highlights· significance 8+

A Mid-Point Review of the Stocks I've Covered8/10

The author argues the market undervalues AMD by viewing it as a GPU second source rather than a full-stack player, a narrative shift that drives multiple expansion.

AMDBullishmoatSubstack · damnang·27 days ago

MU4

7
How the Memory Tax Gets Solved

The post uses Micron's own capacity figures to argue the DRAM shortage is a structural, time-inelastic supply bottleneck that persists into 2028-2029 rather than a normal cycle peak.

Analyst DD

The author argues today's DRAM shortage is structural, not cyclical: HBM eats far more wafer capacity than standard DRAM, new fabs don't arrive until 2027 and get allocated to HBM first, and the three makers are deliberately withholding aggressive expansion. This locks in tight supply and pricing power for memory makers like Micron through 2028-2029.

Evidence
  • By Micron's published numbers, HBM consumes wafer capacity against DDR5 at three to one
  • HBM goes from ~8% bit supply share in 2025 to ~13% in 2027, but on a wafer-start basis from 18% to 30%
  • Micron's Idaho fab does not come online until 2027
  • Q1 DRAM contract prices jumped 90-95% QoQ with a further 58-63% rise expected in Q2 (TrendForce)
  • DDR4 became more expensive than DDR5, a price inversion that cannot appear in a normal market
  • Samsung and SK hynix together hold ~70% of global DRAM and have signaled they will not grow capacity aggressively
Catalysts
  • New lines at SK hynix Yongin, Samsung Pyeongtaek/Taylor and Micron Idaho ship first in 2027-2028, allocated to HBM first
  • Analyst consensus puts meaningful easing at 2H 2027 earliest, normalization 2028-2029
Risks
  • Market reads the shortage as a cycle peak that ends when demand cools
  • Easing could come if AI investment bends
  • NAND is also tight now, with Q2 NAND contract prices up 70-75%
MUBullishsupplySubstack · damnang·today
7
Vera Rubin’s DRAM Cut in Half: Is the Memory Cycle Really Over?

NVIDIA's reported halving of Vera Rubin CPU-side DRAM sent Micron shares down, but the author argues a memory-demand collapse read is premature because lower per-system capacity could loosen supply and lift system count.

Analyst DD

NVIDIA is reportedly lowering the SOCAMM standard build from 192GB to 96GB modules, cutting CPU-side DRAM per rack from ~54TB to ~28TB, which dragged Micron down. The author argues this is not necessarily a demand collapse: SOCAMM bit demand is the product of module capacity, modules per system, and systems shipped, so a lower per-system content can be offset if loosened supply lets more systems install. HBM4 capacity holds while bandwidth nearly triples, and Micron is preparing the HBM4 hot lane and SOCAMM2 for Vera Rubin.

Evidence
  • GB300 NVL72 put forward 20.7TB of HBM3e and 576TB/s; Vera Rubin NVL72 puts forward 20.7TB of HBM4 and 1,580TB/s of GPU memory bandwidth
  • SemiAnalysis reported standard build shifts from 192GB to 96GB modules, dropping CPU-side memory per rack from ~54TB to ~28TB
  • Rack cost falls from about $7.6M to $6.8M, lowering TCO per GPU
  • SOCAMM bit demand = module capacity x modules per system x systems shipped
  • Micron referenced HBM4 high-volume production and SOCAMM2 volume shipment for Vera Rubin together
Catalysts
  • How many Vera Rubin systems actually get installed and in what configuration
  • HBM4 high-volume production ramp for Vera Rubin
Risks
  • 96GB modules halve per-system SOCAMM content
  • LPDDR5X chip supply not keeping up with the full standard build
  • If system count does not make up for the per-system cut, memory demand could genuinely weaken
  • HBM4 capacity does not rise much over the prior generation
MUdemandSubstack · damnang·11 days ago
7
Edge AI Investing Guide: Where Capital Goes After Cloud AI

Phone and PC memory capacity is the critical bottleneck for on-device model size, requiring LPDDR6 transition and 24GB configurations.

MUBullishsupplySubstack · damnang·21 days ago
7
MicroLED: The Optical Technology All Three Memory Makers Are Watching

Micron is evaluating MicroLED optical interconnect technology alongside Samsung and SK hynix to solve memory fabric bandwidth and power constraints.

Analyst DD

Micron, alongside Samsung and SK hynix, has invested in and is seriously evaluating MicroLED optical interconnect (Avicena's LightBundle) because its 'wide and slow' parallelization philosophy mirrors HBM's, and optics could free memory from package geometry constraints. The investment is framed as insurance on what comes after HBM rather than a near-term product plan.

Evidence
  • Micron Ventures participated in Avicena's August 2022 $25M Series A alongside Samsung Catalyst Fund
  • All three memory makers (Samsung, SK hynix, Micron) are on Avicena's cap table
  • Per author's industry network, current employees at all three memory makers are very seriously evaluating MicroLED for memory optical interconnect
  • LightBundle's wide parallel-channel approach extends HBM's I/O philosophy (1024 pins, 2048 in HBM4) into the optical domain
  • Optical link with ~10m reach would liberate memory from CoWoS/reticle package limits
Catalysts
  • Avicena eKit launched March 2026 listing die-to-memory (scale-in) as a first-order evaluation target
  • Disaggregated memory fabric architecture for post-HBM5 custom base die era
Risks
  • Optical memory interconnect ships in product no earlier than after HBM5 — a long time axis
  • HBM4/HBM4E roadmap already locked in as copper/TSV extension
  • Memory latency must be won in tens of nanoseconds, a strict requirement
  • Investment is insurance, not a confirmed near-term production plan
MUBullishexecutionSubstack · damnang·15 days ago

ASML1

7
How does China manage to smuggle in something the size of a school bus from $ASML? Is the real question. https://t.co/Yru8VY5hPO

Raises critical questions about China's ability to circumvent export controls on ASML's advanced lithography equipment, potentially undermining the efficacy of Western technological restrictions.

ASMLBearishregulatory@aleabitoreddit·today

TSM3

6
Glass Substrate: The Order the Money Flows In

TSMC's next-generation CoPoS panel-level packaging roadmap is the channel through which glass enters advanced packaging, with a pilot line targeted for completion this year.

Analyst DD

AI accelerator packages have outgrown organic substrates and the lithography reticle limit, forcing a shift to larger square panels (panel-level packaging) and glass materials. TSMC's CoPoS (Chip-on-Panel-on-Substrate) is the vehicle for this transition, with glass currently most solidly positioned on the temporary carrier side and a pilot line on schedule to finish within the year.

Evidence
  • TSMC's current CoWoS binds an NVIDIA GPU and HBM into one package using a silicon interposer
  • Reticle limit is ~26 x 33mm (~858 sq mm); AI accelerators crossed it long ago and stitch multiple dies together
  • TSMC's package area roadmap moves from 5.5x the reticle to 9.5x then 14x; 9.5x is roughly 8,000 sq mm (~9cm per side)
  • Per Ming-Chi Kuo's June industry check, CoPoS uses a 310 x 310mm glass carrier, with a 510 x 515mm glass panel processed into a glass core substrate floated at the volume stage
  • TSMC is moving on schedule to finish the pilot line for next-generation panel packaging within the year
Catalysts
  • Completion of the CoPoS pilot line within the year
  • Transition to panel-level packaging (PLP) raising area utilization above 75 percent
Risks
  • Glass replacement of the silicon interposer classified by Taiwanese industry press as a long-term task
  • Core substrate adoption for CoPoS is closer to industry check and estimate than confirmation; what substrate early volume runs use remains open
  • Glass is brittle, making yield and customer qualification the central hurdle
TSMexecutionSubstack · damnang·5 days ago
6
If CoPoS Arrives, Who Makes Money First

The post centers on TSMC's CoPoS roadmap as the successor to CoWoS, detailing the timeline, pilot site, and the unresolved questions of where glass fits and when mass production lands.

Analyst DD

TSMC is developing CoPoS (Chip on Panel on Substrate) as the next step beyond CoWoS, using square panels instead of round wafers to fit ever-larger AI accelerator chips with less edge waste. The direction is clear but both glass's structural role and the mass-production timeline remain undecided, with the chairman warning there are no shortcuts.

Evidence
  • First panel format is 310mm x 310mm, with a cited roadmap growing through 515mm x 510mm up to 750mm x 620mm
  • A round 12-inch wafer yields only four to seven large packages per sheet, which a square panel improves on
  • Pilot line laid down at TSMC subsidiary VisEra in 2026; press accounts cite small trial production in 2027 and mass production in 2028-2029
  • Chairman C.C. Wei stated mass production takes another two to three years with no shortcuts; some on the ground see 2029-2030
  • Mass-production site is the fourth phase of the AP7 campus in Chiayi, Taiwan, with NVIDIA cited as first customer
  • TSMC validated a 0.8-millimeter glass core substrate with Ibiden and Innolux under a program named 'glass substrate for CoWoS', while saying mass production is still far off
Catalysts
  • VisEra pilot line in 2026
  • Trial production cited for 2027
Risks
  • Whether glass enters as interposer or core substrate is unfixed, and which path reaches mass production first is open
  • Thin glass interposer (~400 micrometers) makes CTE management tricky and warpage worsens as the panel grows, repeatedly cited as the biggest barrier
  • Timeline weighted toward slipping, possibly to 2029-2030
  • Intel and Samsung are pushing competing panel packaging and glass approaches
TSMexecutionSubstack · damnang·3 days ago
6
MicroLED: The Optical Technology All Three Memory Makers Are Watching

TSMC's partnership with Avicena to optimize photodetector arrays on existing CIS processes gives it manufacturing advantage in optical interconnect scaling.

Analyst DD

TSMC is positioning across optical interconnect technologies, partnering with Avicena (April 2025) to use its CMOS image sensor processes for photodetector arrays and investing through VentureTech Alliance. This is a foundry's hedge against silicon photonics CPO, leveraging an existing high-volume CIS ecosystem.

Evidence
  • April 2025 collaboration to use TSMC CIS processes to optimize and mass-produce LightBundle photodetector arrays
  • VentureTech Alliance (TSMC's VC arm) joined Avicena's May 2025 $65M Series B
  • Silicon PD arrays are already mass-produced by the CIS industry at scale of hundreds of millions of pixels
  • Described as a foundry's hedge: not going all-in on silicon photonics CPO while keeping a foot on MicroLED
Catalysts
  • PD array mass-production collaboration repurposing existing CIS portfolio
  • MicroLED-based CPO market projected by TrendForce to form in 2028, reaching $850M by 2030
Risks
  • MicroLED interconnect commercialization path is multi-year and unproven at scale
  • Fiber bundle and packaging cited as the likely production bottleneck
  • Proprietary architecture with no standards ecosystem
TSMBullishexecutionSubstack · damnang·15 days ago

NVDA7

6
Cerebras WSE-3: The Technical Achievement and the Physical Ceiling

WSE-3 achieves materially higher inference decode speed than NVIDIA's current GPUs through on-chip SRAM, presenting competitive pressure in NVIDIA's inference market.

Analyst DD

The author frames NVIDIA's GPU architecture as structurally advantaged over Cerebras's wafer-scale design: hardware support for multiple precisions plus the CUDA software ecosystem lets rapid software innovation convert directly into performance, while Cerebras's fixed FP16 hardware is 'forever chasing one generation behind.' Even where Cerebras wins on raw decode speed, NVIDIA retains workload portability and dominates training and frontier-scale serving.

Evidence
  • NVIDIA published 1,038 tokens/s for DGX B200 on Llama 4 Maverick vs Cerebras's 2,522 tokens/s, but under different conditions (batch size, precision, speculative decoding)
  • HBM bandwidth climbs each generation: H100 ~3.35 TB/s, H200 4.8 TB/s, B200 8 TB/s
  • HBM capacity scaling fast: H200 141GB, B200 ~180GB per GPU, projected past 1TB by 2028
  • NVIDIA added dedicated FP8 Tensor Cores starting with H100 and extended to FP4 with Blackwell, cutting compute and power
  • Major inference frameworks (vLLM, FlashAttention, Triton, CUTLASS, xformers) are all built on CUDA; any model/framework runs on NVIDIA
  • All major frontier labs train on NVIDIA GPUs, Google TPUs, or their own chips; the $20B OpenAI Cerebras contract is inference only
  • On AWS Bedrock, Trainium handles prefill and Cerebras is used only as a decode accelerator
Catalysts
  • NVIDIA already moving to FP4 (Blackwell) while Cerebras would need 2-3 years to add FP8 in WSE-4
  • HBM capacity projected past 1TB by 2028, widening the capacity gap vs Cerebras's ~44GB SRAM
Risks
  • Cerebras sits at the top tier for inference decode speed on public benchmarks (2,522 vs 1,038 tokens/s)
  • Cerebras signed a >$20B multi-year OpenAI compute deal and an AWS Bedrock deployment
  • Benchmark comparison conditions differ, so the speed gap may not be apples-to-apples
NVDABearishmoatSubstack · damnang·19 days ago
5
Edge AI Investing Guide: Where Capital Goes After Cloud AI

Cloud AI's GPU-centered business model doesn't extend to edge devices, where hardware bottlenecks vary fundamentally by form factor.

NVDAexecutionSubstack · damnang·21 days ago
5
Vera Rubin’s DRAM Cut in Half: Is the Memory Cycle Really Over?

NVIDIA is reportedly cutting the standard SOCAMM/DRAM configuration of Vera Rubin NVL72 to ship racks sooner amid LPDDR5X supply constraints, concentrating the expensive HBM4 lane on hot inference state.

NVDAexecutionSubstack · damnang·11 days ago
3
Glass Substrate: The Order the Money Flows In

NVIDIA's GPU-plus-HBM packages drive the CoWoS demand pushing packaging toward glass, and NVIDIA executives reportedly visited Nittobo to secure T-glass volume.

NVDAsupplySubstack · damnang·5 days ago
3
If CoPoS Arrives, Who Makes Money First

NVIDIA is cited as the first customer for TSMC's CoPoS mass production, with the Rubin Ultra generation expected to use a mix of CoWoS and CoPoS.

NVDAsupplySubstack · damnang·3 days ago
3
Catching Both Copper and Optics: ALAB and CRDO

The post notes NVIDIA's NVLink has effectively monopolized the scale-up interconnect domain, which ALAB is now challenging via open standards.

NVDAmoatSubstack · damnang·9 days ago
3
How the Memory Tax Gets Solved

Nvidia appears only as a section header claiming it seeks to control the entire memory hierarchy as an interface, with the supporting body truncated.

NVDAmoatSubstack · damnang·today

AMD2

6
How the Memory Tax Gets Solved

AMD's acquisition of software startup MEXT is presented as a strategic move to route around the DRAM bottleneck by letting flash behave like DRAM, signaling DRAM-light architecture intent.

Analyst DD

The author frames AMD's purchase of MEXT as evidence that leading chip buyers have stopped waiting for memory prices to fall and are instead building 'detour' paths that use less DRAM. Buying a software prediction layer rather than a chip or memory part signals where the memory market is heading over the next few years.

Evidence
  • AMD bought MEXT, an Israeli software startup, this week
  • MEXT builds a prediction layer that lets flash behave like DRAM
  • MEXT is neither an AI chip nor a memory maker
Catalysts
  • The MEXT acquisition itself as a template for DRAM-reduction strategy
Risks
  • Piece is explicitly informational and not an investment recommendation
  • Each detour technology has points where it breaks, per the author's framing
AMDexecutionSubstack · damnang·today
4
Glass Substrate: The Order the Money Flows In

AMD is named as the qualification customer for Absolics' glass core substrate at its Georgia fab, indicating AMD is an early adopter in the glass packaging supply chain.

AMDsupplySubstack · damnang·5 days ago

MSFT1

4
MicroLED: The Optical Technology All Three Memory Makers Are Watching

Microsoft has entered the MicroLED optical interconnect field as part of broader datacenter bottleneck-solving strategy.

MSFTBullishSubstack · damnang·15 days ago