Semiotic Ghostt retweeted
$AVGO strong positive read-through from UBS TPU checks:
UBS says $AVGO remains deeply engaged beyond v8i, with a new 3nm multi-chip design expected to drive most 2028 volume, followed by a 2nm v10 that could take the lion’s share of Google TPU volume in 2029.
v10 seen using up to 16 dies, AVGO 3.5D packaging and 448G SerDes.
Pretty direct pushback on the “AVGO loses $GOOGL after v8” bear case.
Best pirates sailed the roughest tides... Hold your horses people. The roughness of the open tells me we might've seen the worst already for the week.
Let's unpack the elephant in the room.
MRVL is NOT competing for the logic processor, aka ASIC, in TPU. Also, NOT working on A TPU. MRVL, with this deal, will supply their specially to TPU ecosystem as ATTACHMENTS for inference. What is that Semiotic Ghostt you ask?
Memory controllers, CXL, inference accelerators (logic processor in INFERENCE versions of TPU, like MediaTek, low margin, low cost. AVGO still provides the costly architecture for TRAINING TPU, which is the bulk of BOM, and high margin part), network interface controllers (NICs), near-memory compute, etc.
So, GOOGL is ENCHANCING their TPU ecosystem, which they built with AVGO ever since 2015, and will continue to build with AVGO until at least 2031, where they will presumably sign an agreement again if they don't want their supply taken from them. MRVL, additionally, offers extra CoWoS allocation, which is good for a company hoarding AI infra and designing custom infra to sell to third parties, so does AMD (rumor and sell off a couple days ago).
For this, GOOGL purchased a series of warrants which are conditional on pruchase targets. This is BIG news for MRVL, NO NEWS FOR AVGO. IF YOU KEEP SELLING AVGO LIKE MONKEYS YOU ARE GOING TO BE PISSED OFF WHEN THE 10GW+ PIPELINE GETS UPDATED IN 10DAYS.
If you don't feel ridiculous already, I'll serve you a little example: this is equivalent to assuming the bakery fired the baker because they hired someone else to put the icing on the cake.
NOT FINANCIAL ADVICE DYOR
One of the real ones
A few thoughts on today’s SemiAnalysis TPU report.
1. I am deeply disappointed with the way SemiAnalysis continues to report on these issues. I have raised the same concern before with Micron, CPO delays, SOCAMM, and 8 high HBM4. I do not think they simply fabricate stories out of thin air. The problem is how they communicate them. They often seem too quick to report information that has not yet been confirmed, and even when the underlying facts are directionally correct, the ambiguity in the wording and framing repeatedly creates major misunderstandings in the market.
2. This TPU report is another example. If you actually read the substance, the 2026 AVGO TPU downward revision is explicitly attributed to supply constraints and challenges ramping CoWoS S. The 2027 revision is attributed to greater Broadcom capacity allocation toward other customers such as Meta and OpenAI. The AMD discussion, meanwhile, is about a potential TPU v10 project.
The report does not establish any causal relationship between these two issues. Yet by placing them directly next to each other, it created a framing that could very easily be interpreted by institutional and professional investors quickly scanning the report as a simple
"AMD Long, AVGO Short narrative"
I am not saying that SemiAnalysis intentionally framed it this way. However, if they did not anticipate that kind of market interpretation, then I think this reflects a serious communication and editorial problem. On the other hand, if they did anticipate it, then I have to ask why they did not include even a single clear sentence explaining that the AVGO TPU downward revision was not being attributed to AMD.
3. Anyone who understands the ASIC business even at a basic level should know how dangerous this kind of overinterpretation can be. AVGO signed a long term agreement with Google just this April to develop and supply future generations of TPUs. The claim that AMD is already taking Broadcom’s TPU volume is simply not supported by the information presented in this report.
And if the AMD collaboration they are referring to is related to on package CPU cores, then that is not really a threatening area for AVGO in the first place, because Broadcom is not competing in the x86 CPU market. A much simpler interpretation would be that Google may need x86 cores for specific use cases, such as RL workloads, and AMD could be relevant there.
4. Today I spoke not only with people in the semiconductor industry, but also with a friend who works at a New York hedge fund. He agreed with my view completely, while also pointing out that we cannot ignore either the influence of SemiAnalysis reports or the direction in which capital ultimately moves.
Is this really healthy market behavior? This is the part that concerns me most.
I assume SemiAnalysis understands very well how much influence its reports can have on the market. If so, I believe that influence should come with a much greater degree of care in how information is framed and communicated. This is not the first time this year that I have had concerns about this type of reporting, which is why I felt it was important to call it out directly this time.
5. In the end, I believe the ASIC business comes down to the quality of the IP and the ability to secure sufficient manufacturing capacity. AVGO has competed with Marvell on these fronts for many years and has continued to survive and win. It will continue to face challenges from other competitors on these same fronts, whether that competitor is AMD, Marvell, or someone else.
So let us think again about what we should actually be asking.
Why did AVGO recently enter into such a large strategic MOU with Samsung?
What is AVGO preparing for next in its ASIC business?
From a technical perspective, what are the actual risks facing its ASIC business?
From AVGO’s perspective, in what ways could a more diversified customer base actually become a long term positive?
I believe these are the questions that deserve much deeper analysis.
Why does SemiAnalysis not address them?
I am genuinely disappointed with this report. At this point, I am honestly irritated that I keep having to spend my valuable time rechecking the same facts every time this kind of ambiguous reporting creates another round of confusion in the market.
Semiotic Ghostt retweeted
What a load of crap...SemiAnalysis continues its run of BS. TBT when Jensen refuted SA's news on delays. I cannot believe people (especially those internal to the industry) find this source to be legitimate.
First off, the fact that they claim "AMD has strong IP especially in advanced packaging..." already tells you what you need to know about how dumb this is. AMD does not have any advanced packaging IP. They don't even have fabs...
SemiAnalysis (and follow-on coverage from Tom’s Hardware, The Information, etc.) already reported to say Google’s TPU v9 Humufish is shifting advanced packaging from TSMC CoWoS to $INTC Intel’s EMIB-T.
This was motivated by CoWoS capacity constraints, cost, and scalability in the first place. It is supposed to be multi-million unit packaging capacity starting late 2027 into 2028.
This is why you’ve also seen the news about $SKHY SK hynix testing EMIB for HBM integration and other hyperscalers evaluating Intel packaging.
Near-term (v9 / Humufish) packaging diversification to Intel EMIB-T and further parterships, I refuse to believe Google would ever have a fabless AMD lend a hand when they already have Intel.
The $AVGO TPU unit cuts mentioned in the same note are still mostly about CoWoS supply / allocation pressure in 2027, which is exactly why Google has been looking at Intel packaging in the first place.
Nothing is officially confirmed by Google, Intel, or AMD on either front, so it’s all supply-chain / analyst chatter. The fact that the market wants to react to this is pretty unbelievable to me. The Intel EMIB-T path for the generation after current shipping TPUs has been the more established "rumor" for a couple of months already, but we all know it is happening.
Mediatek is already a strong partner of Intel's.
Semianalysis: "Market chatter suggests $GOOG is working with $AMD on a TPU project in the v10 generation. AMD’s involvement would be the first real involvement in a custom AI ASIC project, despite having a custom silicon team. AMD has strong IP especially in advanced packaging and SoIC. Additionally, CPU IP could also be a draw given Google and its customers are pushing for TPUs with on-package CPU cores for RL workloads."
$AVGO: "We are seeing lower TPU output in 2H26 than previous expectations...For 2027, we also see lower TPU output than our previous expectation of 6mn units of TPU 7 and TPU8i. This is driven by greater allocation towards other customers."
Semiotic Ghostt retweeted
A very material catalyst came about for Western supply chains from $AAOI, $SIVE / Jabil, to $LITE and Coherent:
The Trump administration is drafting a ban of new Chinese optical transceivers and DC devices.
Which would likely hit China’s Innolight, Eoptolink, and other Chinese optical interconnect supply chains.
"The agency would ban all imports of new transceiver models and then exempt many non-Chinese suppliers from the restrictions"
We'll likely see a larger bifurcation of supply chains with even greater importance put on Western players.
Weird effect from the washout… all the TA experts have emerged from their caves claiming to have called top for $SNDK & $MU, and now insisting they know where the precise bottom is… R^2 of this bullshit with a bottom must be 1. Gues we’ll see. Let the hunger games begin. Good luck you deranged bastards.
Well that sucked. Anyway, back from my stay at the psychiatrist
Looking like a bottom here. Market open very much under Max Pain, now up over halfway looks like Put profit taking. Thinking this is the bottom boys. Risk Reversal time? Hold onto your horses... Hopefully some encouraging macro wind on the weekend and July will just be a bad dream.
Also, for anyone thinking otherwise, open-source release of Kimi K3 is net positive for AI trade, bad for sociopaths at Ant/OAI. At the end of the day... don't see a future were models aren't open-source or at least blended. No one will want to sacrifice alpha/metadata for free just to have their business stolen.
That's all for today Folks!
Semiotic Ghostt retweeted
Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI capex.
The fact that spot prices for GPU rentals are at least 2x higher than contracted rates is the missing piece from the discussion about hyperscaler credit, which is the only fundamental factor behind this selloff. Multiple private companies are planning on spending at least 2x more per GPU for compute as contracts roll-off and some have spoken about this publicly.
As contracts roll-off, hyperscale growth rates are going to continue to accelerate as their installed bases of compute reprice higher. Hyperscale operating cash flow growth using a mix of estimates and actuals is modeled to accelerate from 31% in the first quarter of 2026 to 50% in the second quarter. This acceleration should continue for the rest of the year and this is not in estimates which incorrectly model a deceleration in the third quarter from my perspective.
Some math. Consensus estimates are probably for 25-35 gigawatts added by hyperscale and neoclouds in CY28 (using a range as standing up datacenters is hard and a lot of the neos plus labs are still private). At 60b per gigawatt, that is 1.5 to 2.2 trillion in capex. Consensus estimates for hyperscale/neo operating cash flow is 1.3 to 1.4 trillion. I think this gets revised up materially as contracts reprice and growth accelerates so the 100b to 700b that would hypothetically need to be plugged by debt goes away. And their credit profiles materially improve. Not to mention the said 100b to 700b would be less than 1 turn of incremental leverage on consensus EBITDA estimates. And obviously the Nvidia and Broadcom “credit wrappers” help improve creditworthiness as well given their FCF profiles.
OpenAI, Cursor/Grok and the various Open Source inference clouds have accelerated materially over the last two months per public data and Anthropic continues to grow insanely fast while likely generating FCF. This - along with the fact that spot prices for GPU rentals are so far ahead of contract - are the missing pieces from the BofA chart on hyperscale FCF vs. semiconductor FCF.
Hyperscalers are underearning and anyone who signed a contract for GPU compute in 2024 and 2025 is overearning. Operating cash flow will be enough to fund capex but as contracts reprice and cloud growth continues to accelerate then spreads likely come in as well.
Would also note that CDS markets are easy to manipulate - was a huge feature of the GFC - short the stock and then buy the CDS. So I would not put attach much signal to CDS.
Net, net I’m not that concerned about the widening spreads in hyperscale credit. The real risk is that bringing power online and energizing all these GPUs is really hard but we are getting better at this every day.
🤖 Made with AI
Semiotic Ghostt retweeted
translation for retail regard. memory go down because china make model use less memory. white finance man panic because read headline only. asian chad read paper. goal no use less memory. goal maximize cache reuse by orchestrating kv cache across vram, dram and ssd with inter-node transfer so gpu stay busy. white man still no understand.
Looking like a bottom here. Market open very much under Max Pain, now up over halfway looks like Put profit taking. Thinking this is the bottom boys. Risk Reversal time? Hold onto your horses... Hopefully some encouraging macro wind on the weekend and July will just be a bad dream.
Also, for anyone thinking otherwise, open-source release of Kimi K3 is net positive for AI trade, bad for sociopaths at Ant/OAI. At the end of the day... don't see a future were models aren't open-source or at least blended. No one will want to sacrifice alpha/metadata for free just to have their business stolen.
That's all for today Folks!
$AVGO bouncing off the 200SMA. Compelling setup, but astrology aside... Lot of overlooking on the gritty stuff after Hock refused to update +100bn rev forecast, which was honestly very predictable...
Yes, +80% custom ASIC market share. Yes, best-in-class DC switching/routing. The NVDA tax won't hold there forever. But nobody's talking about photonics.
"CPO" gets thrown around constantly. Almost no one is pointing to who actually has product. Broadcom's Tomahawk 6 Davisson... 102.4 Tbps, TSMC COUPE 3D stacking, open laser ecosystem via external ELSFP modules ($COHR mostly, so that's a small risk)... 3rd-gen CPO architecture. Production volume shipping since March.
Spectrum-X Photonics is still ramping, some hard claims switches are having trouble...
Hypers don't like being strongarmmed into single vendor setups, and clearly won't be bottlenecked (insert whole fuss about MediaTek Humufish here)... Don't think Tier 2 will take it for long either...
There could be some networking market share transfer going on sometime soon.