Weekend Sep 19: Fed week ~flat; OpenAI −$278B FCF

The lead

S&P ~flat on the week (−0.08% Fri–Fri) — closed Friday 7,650.50 (+0.17% session); Nasdaq +0.72% week / +0.39% Fri; Dow −1.7% week / −0.18% Fri (Yahoo via Friday settle; Seeking Alpha week wrap).

• Fed week, not a soft landing week. Wednesday’s unanimous 25bp hike to 3.75%–4.00% (first in three years) plus Warsh’s inflation tone left October still live — CME FedWatch ~55% chance of another hike next month (up from ~42% a week earlier). Financials and small caps paid: Russell 2000 −~1.5% week; bank indexes soft.
• 10-year finished ~5.00% after a post-2007 print midweek (~5.04%); Thursday’s sub-5% digests faded into Friday. Rates, not AI, set the weekly tape shape.
• Oil still three-digit after three down days. WTI −1.6% Friday to $100.30 / Brent −0.9% to $103.87 as East-West / Oman workaround hopes outweighed fresh Saudi–Houthi strikes — week roughly flat on WTI after Monday’s spike toward $110 Brent.
• Hardware led; breadth did not. SOX +2.8% Friday carried Nasdaq while Dow/small-caps lagged — concentration into chips, not broad risk-on. Asia Saturday session: Kospi ~+2.3% / Nikkei ~+1.9% on Samsung/Hynix, with oil still a tax on importers (Asia color).
• Overnight lab/financing tape (post-close → Saturday AM): OpenAI’s planning deck shows −$278B FCF through 2030; Anthropic >$100B ARR with a possible November IPO window; Nscale S-1 ($141M H1 rev / $1.0B loss / $56.4B RPO); Anthropic names Accenture Faculty as first embedded frontier evaluator. Same Friday cash session already cleared CleanSpark’s Meta-tied ~$2.28B HY at >4× book.

Open question: Into month-end, does a pinned ~5% 10y and still-$100 oil keep October hike odds sticky enough to fade SOX leadership — or does the overnight lab financing / IPO tape (OpenAI burn vs Anthropic ARR) re-accelerate the AI-capex bid before earnings season?

Positioning

AI capex durability — financing and burn both printed at decade scale; no top-four guide-down. OpenAI’s July deck (FT via Reuters) assumes ~$856B compute/infra by 2030 against ~$840B cumulative revenue and −$278B FCF — capacity dollars dominate the P&L, they do not taper. Nscale’s S-1 puts $56.4B RPO and 10 GW line-of-sight on a public filing beside Friday’s Meta-tied HY print. Break remains a sequential hyperscaler or lab capacity cut, not a headline about burn.

Enterprise agent stall — eval contract landed; Ghodsi says the bind is context, not IQ. Anthropic’s Accenture Faculty embed (≥$1B each / 5y; Anthropic funding Accenture for now) operationalizes Amodei’s pacing evaluator leg — contested as issuer-funded independence, not as whether Faculty sits inside. On a16z, Databricks’ Ghodsi argues frontier models are already “good enough” for most enterprise work and the missing piece is institutional ontology/Genie context. Theme falsifier is still a Fortune 500 naming a production agent with headcount or dollars — not an evaluator LOI.

Inference margin inversion — diffusion’s claim is parallelism, not yet a lab margin print. Ermon’s No Priors bet is that discrete diffusion turns decode into training-like matmul (~10× vs AR at GPT-2 scale; Mercury vs Haiku/Flash/mini speed tier; OpenCall leaving Cerebras for NVDA GPUs). Anthropic’s >$100B ARR alone does not settle cost-per-token vs list; watch whether speed-tier share and disclosed lab margins move together on the next price cuts.

Top stories

Fed-hike week — S&P ~flat, Nasdaq green, Dow −1.7%. First hike in three years to 3.75%–4.00%; 10y reclaimed ~5%; oil stayed three-digit after a Monday spike. Hardware/tech carried; financials, small caps, and cyclicals did not (Seeking Alpha; Reuters wrap).

OpenAI — −$278B projected FCF 2026–2030. FT via Reuters: revenue $36B→$350B (2026–2030), ~$856B compute/infra vs ~$840B cumulative revenue; $122B March raise at $852B valuation on track to exhaust ~2028. Compute and revenue nearly cancel; the burn is “everything else” (news).

Anthropic — >$100B ARR expected; November IPO window in the wires. NYT via Bloomberg: people familiar say annualized revenue expected above $100B this year on enterprise Claude. Separate WSJ/Reuters lanes point at a possible November listing window — treat as secondhand until a filing. Contrast: Altman has said OpenAI will not list in 2026 (news).

Nscale S-1 — $141M H1 rev, $1.02B loss, $56.4B RPO, ticker NSCL. London neocloud with Nvidia/Dell/Point72 among investors; 25k active GPUs / 461k active-or-contracted; Nvidia lease guarantee up to $860M on Texas DC; one customer >50% H1 revenue (CNBC; news).

Anthropic × Accenture — first embedded frontier evaluator. Faculty gets employee-like access for red-team/alignment; each side expects ≥$1B over five years; Anthropic funds Accenture directly for now; non-exclusive, METR still in dialogue (primary; news).

Non-AI session color still in the week. Friday midday: NFLX −~5% on Wells underweight; XENE crushed on psychiatry pause; NUE/STLD guidance miss; crypto proxies ripped with BTC >$80k — beside the SOX bid (CNBC).

Podcasts & interviews

a16z — Ali Ghodsi (Databricks) — NEUTRAL / asr. Enterprise stall is missing ontology/context, not model IQ; “pacing” is bad PR vs real cyber/security controls; four-condition RSI bar mostly unmet today; Uni Gateway + harness multiplexing already cutting $ cost while tokens rise; board anecdote of a scaled org moving frontier→GLM. Recap · Watch

No Priors — Stefano Ermon / Inception — BULLISH / asr. Diffusion as inference-time parallelism (AR decode still sequential); Mercury claimed on-par with Haiku/Flash/mini-nano class while faster; OpenCall left Cerebras custom silicon for Mercury on NVIDIA GPUs; ~20–30% of OpenRouter-style workloads as latency-bound addressable wedge. Recap · Watch

What to watch

October Fed path vs SOX. FedWatch ~55% for another hike; whether hardware leadership survives another week with the 10y near 5% and oil only marginally softer.

OpenAI financing / Anthropic IPO calendar. Deck implies structural pre-IPO capital need through ~2028; Anthropic November window is still wire-only — filing or denial is the tell (OpenAI; Anthropic).

Nscale price range + next hyperscaler-tied HY. S-1 is filed without range; CleanSpark’s >4× / 8.25% sets the credit bar for the next Meta-/hyperscaler-backed print (Nscale).

East-West / Hormuz durability. Workaround working “so long as Iran allows it” per Friday CNBC/JPMorgan color; Rapidan still flags outage risk through month-end (CNBC oil).

Embedded-eval governance stress-test. Whether Accenture Faculty publishes adversarial findings with real teeth — or the issuer-funded critique sticks — matters for how “pace the frontier” lands with enterprises (eval).