Daily Alpha · X
· Post-Market Alpha · by Buzzberg Research
X weighed Meta's Muse Wonder, CXMT's DRAM plan and optical test bottlenecks, while Gavin Baker argued vague supplier MOUs are not promises.
Themes on this desk
Muse Wonder pricing and competition
Zephyr flagged a quoted report of Muse Wonder rolling out at $4 in and $20 out per million tokens as a price war, while Oguz Erkan expects frontier models to become a commodity market.
Validation bottleneck
Michael Sikand quotes a Celestica VP saying 70% of its engineering resources go to validation.
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Citrini says Meta has a real shot at small-business agents after beta-testing Muse Wonder
Citrini's takeaway from the beta test is a really good balance of capability and execution, and the model is surprisingly decent at finance tasks.
Mr. Wang let us take the model for a spin in the beta test and our takeaway is that Meta
It points to small-business agents as Meta's opening.
Source →Chamath says 8090 Labs found Claude Fable 5.1 scores 41% or 61% depending on the coding harness
8090 Labs ran 7 models through 5 harnesses; Claude Fable 5.1 scored 41% in Claude Code and 61% in OpenAI's Codex, and GLM cost $0.93 per success in Codex versus $0.43 in Pi.
Model Benchmarks Are Worthless How can a model score 24% and 59% on the same 113 coding tasks?
Chamath says ranking improvements don't show up in the actual work; the test measured model and harness together.
Source →Gavin Baker says vague AI supplier MOUs are not promises; only contracts count
Baker lists LTAs, SCAs, take-or-pay contracts and fully funded datacenter projects as promises, and says vague MOUs like many signed last year with Asian suppliers are neither contracts nor promises.
Ed, in what sense is a vague MOU with an endless number of outs for both parties a “promise?” Examples
He says those MOUs will be taken seriously only if they turn into contracts.
Watch Whether last year's MOUs with Asian suppliers turn into contracts.
Source →PhotonCap says CXMT's 4F² plan is a company plan, not proof of yields or cost leadership
CXMT reportedly plans a 4F² server-memory product by end-2026; at the same F, 4F² implies roughly 33% less cell area and 50% higher theoretical cell density than 6F², not equal gains in chip capacity or cost.
> China could meaningfully pressure conventional DRAM pricing and profitability.
China could pressure conventional DRAM pricing and profitability, but the announcement does not show CXMT overtaking the big three or dominating HBM.
Watch Verification of CXMT's high-volume yields or cost advantage.
Source →Kawz says Celestica data shows transceiver test windows shrinking from 36 to 18 months
Per Celestica's numbers, testing windows fell from 36 months at 400G to 24 months at 800G and 18 months at 1.6T, while GPU architectures refresh every 12 months.
Recent data at the OCP Summit confirms what has been clear for a while...
Ramping laser fab capacity does not solve the problem if transceivers cannot be validated and qualified fast enough.
Source →