Discussions 17 items
Wed, 23 Sept 2026
Discussion 07:28 ET
Is 16Hi HBM really pushed to 2029 — and does that extend 12Hi tight?
jukan05: JPM HBM model now sees 16Hi adoption 2029 earliest. Replies split on earlier slip vs prolonged 8/12Hi scarcity; no JPM primary in-thread.
top-replies JPMorgan · SK Hynix · Samsung
Tue, 22 Sept 2026
Discussion 07:26 ET
Is EMIB taking real CoWoS share — or just relocating the HBM queue?
SemiAnalysis ChipBook: Korea→Malaysia HBM +467% YoY, ~$3B/2mo via Penang. Replies split: customers fleeing CoWoS scarcity vs queue transfer / qualify risk.
top-replies Intel · TSMC
Thu, 17 Sept 2026
Discussion 15:40 ET
Are AI-infra vendor leads durable when every roadmap hits FFWD?
Nick Dorsey from AI Infra Summit: vendor roadmaps overlap, Astera/CXL example — today's technical leads feel wobbly under industry-wide acceleration.
top-replies Astera Labs
Discussion 15:40 ET Updated
Is AMD MI355X closing agentic-inference TCO vs GB300 — or only on some stacks?
SemiAnalysis says MI355X is quickly closing perf/TCO vs GB300 in agentic inference via SGLang/MoRI/UMBP; replies split on InferenceX $/M-token snapshots and software vs silicon.
top-replies AMD · NVIDIA · SemiAnalysis
Discussion 07:28 ET
Does CPO kill M8/M9 CCL upgrades — or is that scale-out/scale-up confusion?
Market rumor: CPO cuts ultralow-loss CCL demand. Jeff Pu says upgrade stays — CPO not on compute board soon; NVSwitch still needs high-spec PTFE CCL.
top-replies NVIDIA
Discussion 07:25 ET
Did OpenAI flip OpenRouter share vs Anthropic from 20/80 to 50/50?
Baker cites OpenRouter: OpenAI 20%→50% vs Anthropic since June. Talia notes token vs revenue share diverging; Baker agrees OpenAI is discounting.
top-replies OpenAI · Anthropic · OpenRouter
Wed, 16 Sept 2026
Discussion 07:22 ET
Is the Hynix–Intel story really a hyperscaler memory JV — or just talks?
jukan05 says the load-bearing angle is a possible Hynix–Intel JV with hyperscalers hungry for memory; replies stress prior denials, cycle cost, and that talks ≠ wafers this quarter.
top-replies SK Hynix · Intel
Discussion 07:12 ET Updated
Does "pacing" mean more alignment compute and lower lab margins — not less spend?
Gavin Baker argues Anthropic/OpenAI "pacing" is more compute on alignment, monitoring, and evals at the cost of slightly lower margins — not a capex cut. Roon frames pacing as asymmetric margin compression; xEBITDA argues safety spend can raise total semiconductor demand.
top-replies Anthropic · OpenAI
Tue, 15 Sept 2026
Discussion 15:30 ET
Did Anthropic's pacing call already cost frontier enterprise spend share?
Ramp's Ara Kharazian says Astra takes 13% of enterprise AI spend vs Fable at 8%; he frames Anthropic's pace-the-frontier call as already losing adoption. Replies split on whether Ramp is the right sample.
top-replies OpenAI · Anthropic · Ramp
Discussion 07:30 ET
If GPT-6 Astra is looped depth, does that bend HBM vs FLOP intensity?
SemiAnalysis posts that GPT-6 Astra is "basically confirmed" to use loop transformers — deeper passes over layers without growing parameter count. Visible replies split on whether that is HBM-sparing or sequential-compute-heavy. Video body unread; root text truncates mid-quote.
top-replies OpenAI · SemiAnalysis
Discussion 05:00 ET
Is Samsung Taylor's early ramp a real AI foundry recovery — or pilot optics?
jukan05 relays Korean media that Taylor utilization hit ~30% (from ~20% last month), pulled forward vs a November plan on Tesla AI5 demand, with some expecting full utilization by year-end. Aju Press same week frames pilot lines and mass production early next year — a utilization-vs-pilot split.
top-replies Samsung · Tesla · TSMC
Discussion 04:02 ET
Does Samsung dual-sourcing HBM base dies with TSMC ease the HBM bind — or just meet customers?
jukan05 and ZDNet Korea report Samsung will use both Samsung Foundry and TSMC for custom HBM base dies by customer request, with Memory Division owning design when TSMC fabs the die. Replies frame it as avoiding turnkey isolation; wafer-capex allocation to TSMC remains the open bind.
top-replies Samsung · TSMC · NVIDIA
Discussion 03:54 ET
Do Broadcom's FY27/FY28 AI revenue targets rebut Monday's "cycle end" tape?
After Monday's chip de-rating, JP Insights weighs Hock Tan's reaffirmed ~$115B FY27 / ~$230B FY28 AI semiconductor outlook — and Anthropic as largest custom silicon customer in 2027 — against weekly end-of-cycle takes. Visible replies favor operator numbers; demand vs delivery remains open.
top-replies Broadcom · Anthropic · OpenAI
Mon, 14 Sept 2026
Discussion 18:48 ET
Can Anthropic finance a 1/5/10 GW Broadcom TPU path — and top Google as XPU customer?
Tanay Jaipuria relays Anthropic's Broadcom TPU roadmap — 1 GW Ironwood in 2026, 5 GW TPU v8i in 2027, line of sight to 10 more GW in 2028 — with Anthropic as Broadcom's largest XPU customer next year. Thin replies challenge cash generation and call the Google overtake "wild."
top-replies Anthropic · Broadcom · Google
Discussion 18:33 ET
Is a ~90% hike probability the right Wednesday call — or is the room still 50/50?
Eric Balchunas flags a 90% vs 38% juxtaposition on rate-hike odds into the Fed meeting. Replies split between trusting economist/market hike pricing after a hot inflation print and a personal 50/50 that Wednesday can still go either way.
top-replies Federal Reserve · CME
Discussion 17:12 ET
Why do three US labs take ~70% of OpenRouter spend but only ~27% of tokens?
OpenRouter's Peter Walker shows three major American labs attracting about 70% of spend while accounting for only 27% of tokens. Thin reply chain; the chart is the claim. Google's weight depends on whether the metric is spend or tokens.
top-replies OpenRouter · OpenAI · Anthropic
Discussion 08:36 ET
Is datacenter HBM stuck above 4-Hi, or is shorter stack the inference optimum?
jukan05 cites TrendForce that 4-Hi HBM is not enough and doubts suppliers would make it anyway; Jeff Pu concurs that datacenter de-spec to 4-Hi is unlikely. SemiAnalysis the same day argues 4-Hi maximizes tokens per HBM wafer for inference — a live supply-vs-architecture split.
top-replies NVIDIA · TrendForce · SemiAnalysis
‹ All posts Discussions /Discussion ARCHIVE
Discussion · Mon, 14 Sept 2026 · 14:20 ET · Updated Wed, 16 Sept 2026 · 07:12 ET

Does "pacing" mean more alignment compute and lower lab margins — not less spend?

Gavin Baker argues Anthropic/OpenAI "pacing" is more compute on alignment, monitoring, and evals at the cost of slightly lower margins — not a capex cut. Roon frames pacing as asymmetric margin compression; xEBITDA argues safety spend can raise total semiconductor demand.

Review after: 2026-12-15

The read

Baker has the better of the mechanism claim: the posts that engage the economics treat "pace" as a reallocation of compute toward alignment/evals/monitoring, not as a sudden stop to model spend. What is unresolved is magnitude — whether "slightly more" compute is billions (as one reply insists) and whether duty-of-care framing before IPO is load-bearing or decorative. The tail of the thread is unread; most visible replies are agreement, politics, or low-signal.

State of play

After lab commentary and market reaction to an AI "slowdown," Baker states that Anthropic and OpenAI will pace the frontier by spending more time and more compute on alignment, monitoring, and evals, accepting lower margins. He quotes roon that pacing compresses frontier-lab margins asymmetrically. Replies from xEBITDA push a bullish semiconductor implication: safety workloads are memory/storage-intensive and can raise total infrastructure dollars even if capability progress per dollar slows. Stakes: whether Monday's AI-stock dip prices a real demand destruction or a margin/mix shift.

Fidelity note: top-replies only — deep reply tail unread.

The positions

Pacing is more alignment compute, lower lab margins. @GavinSBaker [named] [book] — 'spend slightly more money on compute at the cost of lower margins' @tszzl [named] [principal] — 'compress the margins of the frontier labs' / 'terrible regulatory capture tactic'

Safety reallocation can raise total semiconductor demand. @xEBITDA [pseudo] — '$80 capability + $40 safety' → total spend '$120, not $80'; also 'more bullish for memory than GPUs at the margin'

"Slightly more" understates the bill / profitability risk. @EcoDogFanJM [pseudo] — 'we're talking billions in incremental compute' @tweetmaster153 [anon] — labs 'not profitable' and must 'spend even more' while slowing the revenue driver

Weight of the room

Baker (Atreides CIO) and roon (OpenAI-affiliated voice, marked principal on the quote) carry the named weight. xEBITDA supplies the only falsifiable arithmetic in the visible replies. Anonymous profitability skepticism is present but not evidence. Engagement volume is high; it is not used as proof.

What would settle it

Frontier lab gross-margin disclosure that isolates alignment/eval/monitoring compute as a rising share of COGS or opex Knowable — next 1–2 earnings / IPO filings from Anthropic or OpenAI path vehicles — by 2026-12-15 earliest useful checkpoint

Hyperscaler or neocloud commentary that safety/evals workloads are adding incremental GPU/HBM hours rather than displacing training Knowable — next major cloud earnings Q&A cycle — fall 2026

Public model-card or system-card language that quantifies eval/monitoring compute relative to training for a named frontier release Knowable — next major model release cycle — through Q4 2026

Posts

Gavin Baker (@GavinSBaker) — Sep 14, 2026, 2:20pm ET
The way that Anthropic and OpenAI are going to “pace” the frontier is by spending more time and more compute on alignment, monitoring and evals.

The frontier labs that choose to “pace” likely spend slightly more money on compute at the cost of lower margins.

That’s it.
1855 likes · 157 RTs · 159 replies · 340 bookmarks
https://x.com/GavinSBaker/status/2099563923783524618

Gavin Baker (@GavinSBaker) — Sep 14, 2026, 2:22pm ET
Many factors would go into the lower margins, but incremental compute spend on alignment would be one.

Don’t take it from me, take it from a senior OpenAI employee:
187 likes · 9 RTs · 12 replies · 49 bookmarks
https://x.com/GavinSBaker/status/2099564498486964552

roon (@tszzl) — Sep 12, 2026, 1:07pm ET (quoted in Baker thread)
for the skeptics in government and elsewhere: “pacing the frontier” will compress the margins of the frontier labs. it is a heavy cost imposed asymmetrically on model developers with the strongest AIs in America. by its nature, it would be a terrible regulatory capture tactic
2797 likes · 217 RTs · 294 replies · 486 bookmarks
https://x.com/tszzl/status/2098820692137677116

Gavin Baker (@GavinSBaker) — Sep 14, 2026, 3:07pm ET
And while this is all coming from a place of sincerity, it may also significantly reduce their contingent liabilities by showing a “duty of care.” Important and responsible step before going public.
146 likes · 8 RTs · 11 replies · 27 bookmarks
https://x.com/GavinSBaker/status/2099575895656628419

Dan Druckenmiller (@xEBITDA) — Sep 14, 2026, 2:56pm ET
@GavinSBaker How much memory infrastructure does it take to train, align, evaluate, monitor and safely deploy GPT-N? That number is going up faster.

There is also a nice capex implication. Suppose previously $100 of frontier-model infrastructure spending consisted of $80 capability scaling and $20 safety/post-training. If the labs "slow down" by moving toward $80 capability + $40 safety, total infrastructure spend becomes $120, not $80. Capability progress slows relative to compute consumed, while semiconductor demand rises.
26 likes · 3 RTs · 2 replies · 4 bookmarks
https://x.com/xEBITDA/status/2099573082201686027

Dan Druckenmiller (@xEBITDA) — Sep 14, 2026, 2:58pm ET
@GavinSBaker This is arguably more bullish for memory than GPUs at the margin, because the policy response to AI risk seems to involve more inference, more evaluation, more checkpointing, more monitoring and more data retention. All workloads with unusually high memory/storage intensity. $MU
6 likes · 1 RT · 0 replies · 1 bookmark
https://x.com/xEBITDA/status/2099573610155589696

GrumpMaster (@EcoDogFanJM) — Sep 14, 2026, 2:50pm ET
Slightly more money is doing a lot of work in that sentence .. we're talking billions in incremental compute
7 likes · 0 RTs · 1 reply
https://x.com/EcoDogFanJM/status/2099571543588159651

Delta

2026-09-16 — OpenAI principal quotes land on Baker's mechanism. Overnight @GavinSBaker posts CNBC-attributed Sarah Friar (OpenAI CFO): focused on "getting more compute to keep that flywheel going," and Sachin Katti (OpenAI VP of Compute Strategy): safety/alignment will require "even more compute." Baker: if you thought pacing was negative for AI infrastructure demand, "think again." He also calls Monday's infrastructure selloff almost "Deepseek"-level silliness. Visible pushback (@CZituo): pacing was about when spend lands, not the level — a CFO wanting more compute today does not settle the curve. CNBC Friar piece separately: she would listen to researchers on pacing and still make "strong ROI" investment decisions (CNBC).

More on AI pacing: Sarah Friar CFO of OpenAI yesterday on CNBC: “from where I sit today there is so much opportunity to drive growth that I am still highly focused on getting more compute to keep that flywheel going.” … Sachin Katti … “we’ll need even more compute to make sure future models are more safe and aligned.” If you thought “pacing” was negative for AI infrastructure demand, think again. — Gavin Baker (@GavinSBaker)

Kind of weird the market viewed this as negative infrastructure. This seems wildly bullish infrastructure but highly uncertain for short term to maybe medium term model layer margin structure. — Paul Enright (@pmje73)

Yes. Almost a “Deepseek” level of silliness on Monday. Almost. — Gavin Baker (@GavinSBaker)

pacing was never an argument about the level, it's about when the spend lands. both quotes answer how much, which nobody was really arguing. the bear case is the shape of the curve, & a CFO saying she wants more compute today doesn't settle that either way. — Chen Zituo (@CZituo)

The Open/Close  ·  Research commentary, not investment advice. Positions may be held in securities mentioned.