Daily Alpha · X
· Post-Market Alpha · by Buzzberg Research
The most-discussed themes were Meta's Muse and platform control, the long-end yield spike, AI credit stress and optical hardware supply; the sharpest measured disagreement was whether falling token prices signal an AI demand stall or a mix shift.
Themes on this desk
Rates and auctions
The 10-year posted its largest one-day yield rise in 17 months, up 15bp, and the 30-year hit 5.367%; the $70bn five-year drew a 15.775% primary dealer takedown with the second-biggest tail on record, and Treasury announced up to $6bn of lon
AI credit stress
Hyperscaler CDS set new all-time wides led by Oracle, Nvidia and Meta, Meta's Hyperion data-center bond fell to 94c, and Anthropic's top 1% of customers reached 46% of spend.
Optical and accelerator supply
Lumentum flagged materially higher Nvidia Spectrum-6 CPO laser demand with NPO larger than CPO, Marvell showed a 102.4 Tbps co-packaged-optics platform, UBS cut Nvidia's CoWoS share forecast to 49% by 2027 with Google and AMD gaining, and a
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Third-party testing shows Fastly handles 29.3% of Meta Muse agent traffic, 100% of browser-navigation tasks
A quoted diligence thread reports 738 controlled tests sending Muse agent traffic to test websites; IP analysis showed Cloudflare 37.0%, Fastly 29.3%, Meta 19.4%, AWS 14.4%. Fastly was used in 100% of tests requiring browser navigation (e.g., 'go to my favorite restaurant's website and find me the menu') and 100% of tests requiring website navigation (e.g., booking a doctor appointment), while Cloudflare was used in 100% of API/curl tests. The poster calls the Fastly role 'greater than analysts have noted.'
the 738 tests, Muse sent traffic from: > -Cloudflare: 37.0% > -Fastly: 29.3% > -Meta: 19.4% > -AWS:
If agentic browsing becomes a meaningful traffic category, Fastly could capture incremental CDN/edge volume that sell-side models may not yet reflect, a potential positive revision driver for FSLY.
Watch Independent replication of the traffic-share split, or Fastly commentary/analyst notes confirming agentic-browsing traffic contribution in reported results.
Source →Lumentum: Nvidia Spectrum-6 CPO UHP demand up materially
Per @aleabitoreddit, Lumentum said in a Stifel meeting at ECOC 2026 that Nvidia Spectrum-6 co-packaged-optics ultra-high-power laser demand increased materially.
Lumentum in a meeting with Stifel at ECOC 2026 stated: - $NVDA Spectrum-6 CPO UHP demands increased materially.
Rising CPO-attached laser demand is a leading indicator for Nvidia's next-gen switch ramp and for the optical component suppliers levered to it, ahead of reported networking revenue.
Watch Nvidia networking/switch revenue acceleration and Lumentum commentary on CPO volumes in subsequent quarters; a demand plateau would falsify the read-through.
Source →Trump Accounts expansion: universal at-birth investment accounts with $250 seed per child
The author celebrates the expansion of Trump Accounts giving every child an investment account at birth, projecting $50k at 18, $175k at 30, and $1M at 55 if started at birth with modest saving. The quoted Vlad Tenev post notes thousands of Indiana kids became investors, with $250 seeded into each of their Trump Accounts courtesy of @altcap and the Brad Gerstner Family Foundation, thanking Brad Gerstner for leading the way.
> > $250 into each of their Trump Accounts, courtesy of @altcap and the Brad Gerstner Family Foundation.
A universal at-birth equity account program would channel a recurring, long-dated retail flow into US equities and asset managers, and the $250-per-child philanthropic seeding creates a template for private capital to scale adoption ahead of any federal rollout.
Watch Whether federal legislation or pilot programs formalize at-birth accounts and whether private foundations continue seeding balances; confirmation would come from enrollment counts and asset-manager disclosures of new small-balance accounts.
Source →TPUv7 benchmarked 56% ahead of Nvidia GB200 NVL72
A cited SemiAnalysis alert claims vLLM maintainers showed TPUv7 reaching 700 tok/s/user, 56% better performance than Nvidia GB200 NVL72 via megakernel optimization on Kimi K3, with the poster calling the TPU software externalization 'full steam ahead' and 'ultra important to follow'.
SHOWN THAT TPUv7 CAN GET 700 tok/s/user, 56% BETTER PERFORMANCE THAN NVIDIA GB200 NVL72 THROUGH MEGAKERNEL OPTIMIZATION ON KIMI K3.
If third-party benchmarks confirm a large TPU inference advantage, it pressures Nvidia's inference share and strengthens Google's internal-silicon cost position versus merchant GPU buyers.
Watch Independent reproduction of the 700 tok/s/user figure and evidence of external TPU software adoption; Nvidia inference benchmarks closing the gap would invalidate.
Source →HOGO collapse closes USGC-to-Rotterdam diesel arbitrage
BurggrabenH reports HOGO collapsing from 24.69 cpg to 4.43 cpg, shrinking the USGC-Rotterdam differential from $10.37 to $1.86/bbl, which is insufficient to pay for a clean tanker, closing the arb and implying the market is already pricing a major impairment of the US-to-Europe diesel outlet.
Scary So HOGO collapsing from 24.69 cpg to 4.43 cpg means the diff collapsed from $10.37 to $1.86/bbl which is
A closed diesel export arb pressures US Gulf refining margins and clean tanker demand on the transatlantic route, and signals the market is front-running a possible US diesel export ban.
Watch Whether the HOGO spread stays below the ~$2/bbl freight breakeven and whether a formal US diesel export ban is announced.
Source →