The Investor

The consideration tells you more than the valuation

SpaceX did not write a check for Cursor, which is the part worth sitting with. It issued roughly 389 million Class A shares, closing one of the largest venture-backed takeovers on record entirely in its own privately marked paper, per TheSequence. The Information reports the deal is done, with the Cursor brand set to be phased out on some products. For Anysphere's holders, that converts a cash exit into a position in an unlisted acquirer — or rather, into a position plus a negotiation about when they are allowed to leave it. Proceeds modelled at a 30-50% haircut to SpaceX's own mark read very differently from the $60B headline, and registration rights, collars and lockups become the real negotiation. The strategic logic is distribution, not revenue: Grok 4.6 is routed natively into Cursor, Grok Build and GitHub Copilot.

Nvidia's 25% is a ceiling, not a commitment

Nvidia is close to providing roughly $100B of credit support for OpenAI to lease an Ohio data center, negotiated down from a reported $250B, The Information reports. It covers phase one only, with later phases deferrable for a couple of years. Nvidia wants exposure capped at 25% of total financing - the identical ceiling on Monday's $500B syndicate with six of Wall Street's largest firms. Run the arithmetic the announcement omits, because someone should: if $100B is 25%, phase one of one project implies about $400B of total financing, leaving roughly $300B for non-Nvidia capital. Nvidia holds $80B+ in cash and securities and generated $49B in a single quarter, so this is an optics constraint on equity and rating perception, not a funding constraint. Consequence for origination: the marginal AI infrastructure dollar is now a credit dollar. That is why Nvidia is discussing $3B into SB Energy's IPO and investing in Lancium behind Stargate Texas while SoftBank commits $64B+ to OpenAI. Four separately-branded deals can share one point of failure.

Primary demand accelerated in the same five days

Cognition is in early talks for more than $1B at $40B, under three months after a $26B round, on roughly $1B annualized revenue - about 40x, a 54% step-up in a quarter. Lovable's $400M Series C doubled its December mark to $13.3B against a ~$600M ARR trajectory, near 22x. Databricks targeted $1B, saw $15B of investor interest, and took $5B at $190B on a $7B run rate growing over 80% year over year, near 27x and the best fundamentals on the board.

Where the two readings diverge

TheSequence treats Cursor as a distribution asset paid for in paper, and reads Thrive's sale as the informed holder stepping toward the exit. The Information reads the same transaction as evidence that industrial and vertically integrated acquirers are now top-of-market bidders for AI developer tools - which widens the exit set rather than signalling a top. Both can hold, and this is probably the more useful way to carry them. What is not in dispute is the structure: at these prices, the buyers with a choice are spending equity and capped credit, not cash.

At current marks, the parties best positioned to price AI assets are the ones handing the valuation risk to someone else.

Stated intent versus a realized line item

UBS's Karl Keirstead says blue-chip executives are "absolutely articulating a view" that software vendor spend should come down 30% over the next three years, on the theory that models are now good enough to make building it yourself a real option, and his own forecast is "a rocky ride for the next 12 months." That is a TAM-contraction statement coming from buyers rather than a valuation opinion coming from the sell side, which makes it more interesting and, unfortunately, no less stated intent. Enterprises announce vendor consolidation every cycle and mostly fail at it, usually quietly. The reason to weight this instance differently: IBM's mainframe purchases actually declined as customers moved budget to AI, and the shares crashed. Intent has started converting into a reported financial. The exposure sits in the pricing model rather than the product, and if revenue scales with headcount while AI compresses headcount, seat-based ARR is the line item at risk.

There is no software sector, only business models

The recent prints produced dispersion, not a sector read. Palantir was rewarded for continued scorching revenue gains, Atlassian soared on cloud growth, and Anthropic told investors revenue grew roughly 14x year over year in Q2 — off an undisclosed base, which is the part of that sentence doing the real work. In the same window IBM crashed and Workday sits down 7% year to date. Index-level exposure is the worst available expression of a view on software right now, because the dividing line runs through the revenue model rather than the vertical.

Why the screen exists precisely because the price does not

Private equity has not been sitting out software for lack of dry powder. Their own private software portfolios carry the same AI exposure as the public names they would be buying, which is an awkward thing to explain twice in one investment committee. That is why a completed take-private would signal more than one deal: it would mean PE has rebuilt underwriting conviction on AI-resilient software cash flows. Meanwhile KeyBanc's Ader attributes the bounce to money rotating out of high-momentum names into "relative losers of late" — technical flow rather than fundamental re-rating — and the giveback inside 24 hours supports him. Marking private SaaS up on that flow is LP-credibility risk with no offsetting upside. Note also the spread being underwritten across: private application-layer AI assets are being marked at 20-40x forward revenue while public seat-based incumbents screen on free cash flow.

What resolves it, and when

Salesforce reports in roughly two weeks against consensus of more than 10% growth to about $11B, with AI product traction the watched line, and it is fair to treat that as a scheduled binary for the seat-based cohort. On the take-private side, the decisive detail is not whether a price prints but what it is priced on. A revenue multiple means PE is still paying for growth in seat-based software and the de-rating thesis is overdone. A cash-flow multiple confirms the thesis. If the talks collapse, the complex loses its reference comp for another two quarters while fundamentals keep eroding.

The floor under software is not a chart pattern. It is a take-private multiple that nobody has published yet.

The arithmetic that breaks per-prompt routing

The underwriting question is never whether one clever routing example saves money; it is what the general rule does to the invoice. Staying on a warm model costs ten percent of the base rate on reused history plus the full rate on new tokens, while switching bills the entire prefix cold on the new model, and if you set those two equal for the Opus 5 / Haiku 4.5 pair you get a break-even history-to-new-token ratio of 8. With a 200-token instruction, that means the whole conversation has to sit under 1,600 tokens for the switch to pay. A system prompt plus tool schemas clears that before the first user turn. Per-prompt routing inside an agent session is not a marginal trade. It is structurally outside the profitable regime for short instructions.

So the routing signal that survives contact with billing is projected output length, not prompt difficulty. Output savings run about $20 per million tokens, which puts the payback threshold near 1,450+ output tokens per switched turn just to recover the re-prefill penalty. Tool calls and control-flow turns lose money. Full-file writes and large diffs win. Post-compaction turns are free switch points, because the cache is already invalid. The charitable read on how the category missed this is benchmark provenance rather than fraud: routers were first measured on short, independent prompts where re-prefill cost nearly vanishes, and long agent sessions inherited that conclusion without anyone revalidating the conditions.

The counter-signal that points the other way

NVIDIA shipped NeMo Switchyard, an open-source router that sends each step of an agent workflow to the cheapest capable model, and it landed next to Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters and 1M context under the permissive OpenMDW-1.1 license, per TheSequence. DeepSeek-V4-Pro also went generally available with native OpenAI Responses API compatibility, which makes it a drop-in endpoint substitute rather than a migration project. TheSequence reads all of this as a hard ceiling on model-only gross margin, with Grok 4.6's reported $2/$6 per million tokens now functioning as a price cap rather than a price. That is a defensible reading. It is also a reading about a different unit of analysis.

Reconciling them is the actual thesis

Both hold, at different unit boundaries. Choosing a cheap open-weight model for an entire session, or switching where the cache is already dead, captures the commoditization TheSequence describes; switching mid-session on prompt difficulty destroys it. The category therefore splits, and cleanly: pure-play prompt-difficulty routers deserve multiple compression, while cache-boundary schedulers, cross-model cache portability and per-turn cost attribution (the FinOps-for-inference wedge) deserve a premium. Demand is not the uncertainty. A single coding task drags 400K to 2M cumulative input tokens through the meter because every turn re-sends the full transcript, which makes inference a first-order COGS line for every agentic software company.

The defensibility update nobody has priced

This is probably wrong in one direction, but the reclassification worth making is prompt caching from pricing feature to per-session switching cost. It compounds with session length, it cannot be engineered away at the router layer, and it makes frontier-lab retention in agentic workloads stronger than consensus. It matters more now because model quality keeps converging: GLM-5.3 was produced entirely by post-training the same 743B base model as GLM-5.2, meaning capability catch-up no longer requires a pretraining run. When weights converge, retention has to come from somewhere, and the somewhere is the cache. It ages badly if cache portability gets standardized across vendors, if output-heavy workloads grow enough to make switching pay on volume, or if the labs price caching away to buy share. Absent those, two companies with identical ARR and identical model choices post materially different gross margins purely on caching discipline.

Prompt caching is not a discount. It is a switching cost that grows with every turn, and it turns naive model routing from arbitrage into a cost overrun.