MANISH

Kernel architect (ARM/x86) | AI chip design, infra & memory systems | low-level performance & power management

[Sam Altman @sama]

build and publish web apps with chatgpt!

i really wish i had this when i was a kid, but i do miss hypercard.

[SemiAnalysis @SemiAnalysis_]

@Klaytonlolus Where did we say bearish MU? The note on their HBM delay literally said it was good because regular DDR margins are higher than HBM. But you wouldn't know that because you didn't actually read our note.

[パウロ @paurooteri]

HBM >> SOCAMM2

[Jukan @COMPUTEX @jukan05]

NVIDIA COMPLETES APPROVAL OF HBM4 FROM ALL BIG THREE MEMORY MAKERS.

[Jukan @COMPUTEX @jukan05]

Meritz Securities Korea comment on the SoCAMM2 rumor:

Hello, this is Sunwoo Kim and Seungsoo Yang from Meritz Securities.

We are sharing our comment on the capacity adjustment of NVIDIA NVL72 LPDDR5X SoCAMM2 modules.

⸻

SoCAMM2 modules: Total demand is actually increasing

A correction based on misunderstanding is an opportunity

Market concerns are emerging over the reduction in LPDDR5X SoCAMM2 module capacity for NVIDIA NVL72, from 192GB to 96GB.

However, this is merely noise that occasionally occurs when stock prices lead during a DRAM upcycle. Investors should keep in mind that the DRAM upcycle has still not even reached the mid-cycle stage.

We continue to recommend that investors should only start reducing memory exposure from the “shoulder” of the industry cycle. The current level is still merely around the “thigh” of the cycle.

According to our channel checks in Taiwan and Korea, NVIDIA has recently changed the shipment mix of SoCAMM2 modules, while actually increasing total order volumes.

While 192GB modules are being cut by half, 96GB modules are instead increasing by nearly sixfold and becoming the mainstream product. 64GB modules are also increasing by 50%. This appears to be NVIDIA’s pragmatic workaround to prioritize expanding Vera Rubin system shipments amid tight memory supply.

As a result, total SoCAMM2 LPDDR demand is expected to increase by 10–20% compared with the previous forecast.

Moreover, since Vera Rubin systems can be upgraded to higher-capacity modules at any time after initial shipments, this should be interpreted as a positive event that expands the medium- to long-term memory demand ceiling.

For SoCAMM2 module component, materials, and back-end process suppliers, the shift toward higher shipments of lower-capacity products is expected to lead to a sharp upward revision in 2026E/2027E shipment volumes.

⸻

Value chain check

According to our channel checks, there appears to be no component- or materials-level design change to SoCAMM2 modules resulting from this capacity adjustment.

In other words, although memory capacity per module is being lowered, there is no major change in substrate specifications, key material composition, or back-end process complexity. Therefore, the related value chain should fully benefit from the increase in shipment volume.

Rather, in the next-generation products, we expect an increase in substrate layer count, more advanced stack structures, and broader adoption of low-loss materials in order to support high-speed signaling and ensure power integrity.

Therefore, this adjustment should be interpreted as a double positive: in the near term, it expands SoCAMM2 module shipments; and in the medium to long term, it drives higher-value substrate and materials specifications.

[Quiver Quantitative @QuiverQuant]

BREAKING: We just caught some interesting new stock trades.

Representative Josh Gottheimer just filed purchases of:

- SanDisk, $SNDK
- Micron, $MU
- AMD, $AMD
- Palo Alto Networks, $PANW

Gottheimer sits on the House Subcommittee on AI.

Full trade list up on Quiver.

[Micron Technology @MicronTech]

Every search, recommendation, simulation and AI-generated response depends on rapid access to #data.​​

From training large AI models to enabling real-time inference, #DRAM helps keep today's most demanding workloads moving at the speed innovation requires.​​

This National DRAM day, join us in celebrating the technology that helps bridge the gap between processors and data, enabling everything from smartphones and PCs to #AI infrastructure.

[MANISH @OrbitHigher]

@italianinvesto Super confused:

As per their ER, AVGO is growing 30% QoQ ($22 -> $29B) and 60% AI Revenues QoQ ($10B -> $16B)
FY AI Revenue $56B -> 2027 $100B 78% YoY

I guess it just ran up a lot. but Hock mentioned 'substantial' Google revenues multiple times.

[MANISH @OrbitHigher]

@sheslee Amazing call!! Brilliant.

[Eric Daugherty @EricLDaugh]

🚨 WOW! Secretary Rubio is taking NO BS from Dem Rep. Ted Lieu, he keeps DESTROYING them

LIEU: Trump can't even stay awake during meetings!

RUBIO: He's NOT falling asleep

LIEU: YOU'RE LYING

RUBIO: Is this a JOKE!? This is a Foreign Affairs Committee you're asking about SLEEP [...] Absurd and ridiculous. I can't believe we're in a Foreign Affairs Committee with questions about someone who thinks he's a MEDICAL EXPERT when he's NOT! We HAD a cognitively impaired president for 4 years. THIS president is a guy that literally DOESN'T sleep, works long hours every day. I talk to him day and night, HE WORKS INHUMANE HOURS. I've been on trips where he doesn't sleep the whole flight, he's wandering the hallways to wake people up and talk to! He has an INCREDIBLE amount of energy.

He's so good.

[Hardik Shah @AIStockSavvy]

Bought $AVGG ( $AVGO). I think it's overreaction.

[MANISH @OrbitHigher]

@mzuhair123 Pls read the full article

@SemiAnalysis_

[MANISH @OrbitHigher]

@WisemanCap 😂😂

[MANISH @OrbitHigher]

@tuolaji2024 Do you have full article link from @SemiAnalysis_ @dylan522p

[Wall St Engine @wallstengine]

Apple’s overhauled Siri will reportedly use Google’s Gemini models running on NVIDIA Blackwell B200 chips, per The Information.

$AAPL tried running a modified Gemini model on its own Private Cloud Compute servers, but performance was reportedly too slow.

More complex Siri requests may rely on Google’s cloud infrastructure, while Apple is still expected to emphasize on-device AI at WWDC.

[Hardik Shah @AIStockSavvy]

$AVGO | 𝐁𝐍𝐏 𝐏𝐚𝐫𝐢𝐛𝐚𝐬 𝐄𝐱𝐚𝐧𝐞 maintains 𝐎𝐮𝐭𝐩𝐞𝐫𝐟𝐨𝐫𝐦 on 𝐁𝐫𝐨𝐚𝐝𝐜𝐨𝐦, 𝐫𝐚𝐢𝐬𝐞𝐬 𝐏𝐓 𝐭𝐨 $𝟔𝟒𝟎, 'Primed for Growth'

Analyst sees the market's negative reaction to near-term AI sales as shortsighted, citing accelerating AI orders supporting visibility through '28.

[MANISH @OrbitHigher]

@PythiaR Sorry I am honestly confused here.

AVGO guided 80% YoY revenue growth, 35% QoQ and 200% AI Revenue growth, and it is down 15%

Plus $100B AI Revenue in 2027 which is 78% YoY growth.

What am I missing?! Perhaps it had a strong run up to ER.

[MANISH @OrbitHigher]

@BlueChipPremium AVGO guided 80% YoY revenue growth, 35% QoQ and 200% AI Revenue growth, and it is down 15% ... CIEN down 10% is nothing ;-)

[Dylan Patel @dylan522p]

I love that people reposting what we said miss most of what the note says.
Happens all the time.

[Kaushik @WisemanCap]

$AVGO PT Raised to $580 at JPMorgan

We came away from the F2Q (Apr-Q) print incrementally more positive on AVGO’s growth outlook over the next couple of years, reinforcing our conviction that the AI compute buildout remains early-innings and that AVGO is the most strategically entrenched provider of ASIC/XPU compute/networking. Apr-Q results were solid and, despite a fair amount of hand-wringing ahead of the print, Jul-Q guidance still landed ahead of Street consensus with revenue of $29.4B (+84% Y/Y) and AI semis revenue up 200%+ Y/Y to $16B. But the more important signal was orders: AI bookings eclipsed $30B against $10.8B shipped, lifting backlog sharply and raising revenue visibility even further. Mgmt even struck a very confident tone on FY28 visibility (unusually early), with what we see as ample room for that FY28 pipeline to keep building over the next 12-18 months.

Management detailed an expanding, increasingly diversified customer set, reaffirming a multi-generation TPU engagement with Google, underpinned by substantial revenue commitments from Google and reciprocal supply commitments from AVGO (which we have repeatedly highlighted - link, link); Anthropic, layering an incremental ~5 GW of next-gen TPU from FY27 atop its >1 GW FY26 footprint; OpenAI, with first silicon already delivered and 1.3 GW slated for FY27 within its 10 GW/2029 commitment; Meta’s multi-generation MTIA program (3 GW through 2028, initial 1 GW XPU-plus-networking order delivering in 2H27); and two further customers ramping from late 2026, with $6B of POs already in hand. Notably, for AVGO’s >20 GW XPV compute platform built with Apollo and Blackstone to fund capacity for Anthropic and OpenAI, we estimate only ~5 GW is delivered in FY27 (~1.3 GW OpenAI + ~3.5 GW Anthropic), implying ~15 GW still to ship in FY28 for those two customers alone; folding in Google, Meta, and other at-scale programs, we think FY28 AI revenue could approach $300B - far in excess of current expectations.