Nscale lines up $3.36B pre-IPO, and NVIDIA is both the anchor and the customer

Good morning 👋 NVIDIA just wrote a check to its own GPU customer, days before that customer starts its IPO roadshow. The filing explains why, and it isn't generosity.

In today's issue:

• 🔭 Nscale lines up $3.36B pre-IPO, and NVIDIA is both the anchor and the customer
• 🔬 Nubank simulates chatbot changes before real customers see them, and two more papers worth your attention
• 🚀 Confido raises $55M to fix CPG reconciliation
• 🛠️ Turn idle GPU cycles into batched robotics simulation with NVIDIA Warp
• 🔗 Cognition's revenue run rate, Ringg's call deflection, invideo's color-grading speedup

🔭 THE ONE THING

🏗️ Nscale lines up $3.36B pre-IPO, and NVIDIA is both the anchor and the customer

Nscale just closed $2.36 billion of a $3.36 billion convertible note round, led by Third Point, with Apollo, Citadel, Hudson Bay, the Abu Dhabi Investment Council and 8090 Industries filling out the syndicate. NVIDIA is in for the other $1 billion, landing mid-November, and every note converts into equity the moment Nscale prices its IPO. That structure tells you what this really is: a bridge loan dressed as a funding round, priced to dodge setting a valuation until public markets do it instead. NVIDIA anchoring its own GPU customer's balance sheet ahead of an IPO is the part worth sitting with, not the topline number. If you're building on neocloud capacity, the takeaway is that the chips are still the scarce asset, and the people selling them are now also underwriting who gets to resell them.

🔬 RESEARCH HIGHLIGHTS

• Nubank tested chatbot changes on synthetic customers before real ones ever saw them. Its highest-volume chat-support agent in Brazil got run through Snowglobe, a simulator where fake customers and fake tool calls stand in for the actual backend. Simulated scores tracked production closely enough across four shipped versions to trust the loop, and simulation-guided iteration lifted transactional NPS by 36.69 points in a live A/B test. They also screened open-weight model swaps across 16,000-plus simulated conversations before picking one for another rollout, which pushed self-service rate up 8.82 points to Nubank's best ever, no NPS hit. In a regulated industry, that's thousands of shots on goal you'd never risk live. Paper

• Visual-text alignment scores are measuring geometry, not understanding. Swap the image tokens in 13 multimodal LLMs for pure noise and task accuracy collapses. Four standard alignment metrics, CKA, SVCCA, MIR, principal-angle cosine, never notice. The culprit: MLP projections shove both modalities toward the same output directions regardless of content, faking the "shared representation space" everyone assumes is happening. If you've cited an alignment score to justify a multimodal architecture choice, this is the noise test you skipped. Paper

• A model's excuse for rejecting a candidate is only half real. Ask an LLM to pick between two profiles and it usually rejects the loser by naming a missing fact. Researchers inserted that exact fact into the profile and re-asked: doing so flipped the answer far more than an irrelevant control sentence (odds ratio 3.57). But moving the same fact to the option the model hadn't named didn't hold up under correction, and an irrelevant sentence swayed the model more just by landing on the named rival than on the ignored option. Where a fact sits moves the answer almost as much as what it says. Credit to the authors: they flag their own parsing bug, present in 17% of responses, that would've cut the headline result from six surviving effects to four. Paper

🚀 AI STARTUPS

• Confido closed a $55M Series B, led by Insight Partners with Y Combinator still on the cap table, to keep building its AI back office for CPG brands. Cash application, deductions, trade promo, all collapsed onto one data layer instead of the usual pile of disconnected spreadsheets. 250+ brands run through it already, including divisions of Unilever, Mars and Nestlé, and Confido claims implementations land two to three times faster than the legacy systems it's replacing. Unglamorous plumbing, but reconciliation is a real tax on anyone moving physical goods. Confido raises $55M Series B

🛠️ TRY THIS

Turn idle GPU cycles into batched robotics simulation with NVIDIA Warp

NVIDIA just anchored a $3.36B neocloud raise on the compute-supply side. On the demand side, its new MjWarp guide shows what you actually do with that GPU capacity: move a MuJoCo robotics task off single-world CPU execution onto thousands of parallel worlds. The code delta is smaller than you'd guess. Worth an afternoon if you're running RL rollouts or domain-randomization sweeps.

1. Get the task running clean in stock MuJoCo on CPU first. That's your ground truth for parity checks later.

2. Port to MjWarp as a single world (nworld=1) and diff the trajectory against the CPU run before touching batch size.

3. Tile your initial state with np.tile(), scale to hundreds or thousands of worlds, and size nconmax/njmax up front instead of debugging silent truncation after the fact.

4. Wrap the step loop in wp.ScopedCapture() for CUDA graph replay, and only start timing after a GPU sync plus a warmup pass. Skip that and you're benchmarking queue latency, not compute.

Prompt: I have a MuJoCo simulation running on CPU (paste my step loop and model-loading code). Help me port it to NVIDIA's MjWarp for batched GPU execution: show me the mjw.put_model/make_data calls, how to tile my initial state across N worlds, and how to wrap the step in a CUDA graph capture.

🔗 QUICK LINKS

• Cognition AI The Devin maker is tracking $1B in annualized revenue this month, roughly double the $492M run rate it reported in May, per Bloomberg. Citi, Goldman, GE Aerospace, and the U.S. Navy are on the customer list now.
• Ringg OpenAI says the voice-agent startup's GPT-5.6 stack now closes 65% of customer calls without a human, at roughly a tenth the cost of running GPT-4.1. Details
• invideo On GPT-6 Astra, the video editor grades color three times faster and turns out 50 custom effects a day. Details

That's the stack worth knowing today. Back tomorrow.

Pradeep Perugu