One-Line Summary

Nvidia ties financing to infrastructure with over $40B in AI equity stakes while rolling out rack-scale and observability features that shift how teams plan capacity and monitor training.

Big Tech

Nvidia unveils Groq 3 LPU racks and Vera CPU systems

Nvidia introduces a dedicated inference rack built around the Groq 3 “language processing unit” and a CPU rack using its new Vera chips to speed agent-like AI tasks and large-context inference. The company says an LPX rack pairs 128 Groq 3 LPUs with its Vera Rubin NVL72 rack, claiming up to 35x higher throughput per megawatt and 10x more revenue opportunity, while the Vera CPU rack aggregates 256 liquid‑cooled Vera chips for agentic workloads. ¹

These launches sit atop Nvidia’s rack-scale NVLink architecture: GB200 NVL72 extends a coherent NVLink domain across a full rack, delivering 1.8 TB/s per GPU (130 TB/s aggregate) and imposing a sharp drop when crossing domains (~50 GB/s over InfiniBand/Ethernet). To place jobs correctly, Slurm’s new topology/block plugin and the --segment option let teams request atomic NVLink-local segments, balancing queue time and performance. ²

Keeping those racks saturated also depends on live diagnostics. Nvidia’s NCCL Inspector now exports real‑time collective metrics in Prometheus mode, making it easier to spot communication bottlenecks; in a large LLM pretrain test, introducing network constraints cut per‑GPU compute from ~310 TFLOPs to ~268 TFLOPs (about 13%), a gap teams could trace on Grafana dashboards. ³

Ecosystem projects are tuning for this shape of workload: one DGX Spark forum post claims an open inference engine (“Atlas”) achieves sub‑2‑minute cold start and 100+ tokens/s on Qwen3.6‑35B FP8 across supported models, underscoring a push toward faster, steadier inference. ⁴

Industry & Biz

Nvidia commits $40B to AI equity stakes in 2026

Nvidia is taking ownership positions across the AI stack, with CNBC tallying more than $40 billion committed so far in 2026 — including a $30 billion investment in OpenAI and agreements to invest up to $3.2 billion in Corning and up to $2.1 billion in data‑center operator IREN. FactSet data also shows Nvidia participating in roughly two dozen private startup rounds this year. ⁵

The bets span optics, compute, and data‑center capacity: the IREN pact includes deploying up to 5 gigawatts of Nvidia DSX‑branded infrastructure, while Corning plans three new U.S. facilities dedicated to Nvidia optical tech. CNBC also notes earlier moves — for example, a $5 billion stake in Intel that swelled above $25 billion — as Nvidia widens beyond chips into the broader AI supply chain. ⁵

Analysts call many of these arrangements “circular” — investing in partners who, in turn, buy Nvidia gear — but argue they can still deepen Nvidia’s moat if executed well. TechCrunch highlights this tension and the potential strategic upside despite concerns. ⁶

At the platform edge, demand signals show up in small ways: on Nvidia’s developer forums, a customer building an autonomous multi‑agent document system asks to raise NIM API limits from 40 to 200 requests per minute — a reminder that throughput planning is now a product requirement, not just an ops detail. ⁷

What This Means for You

Nvidia’s equity push ties component, compute, and capacity closer together. If your roadmap depends on GPU access, expect more bundled offers with optics and data‑center commitments from partners like Corning and IREN — potentially improving availability but concentrating bets on Nvidia‑aligned stacks. Bring procurement and product together on capacity assumptions. ⁵

If your team runs training or fine‑tuning, rack‑scale NVLink changes how jobs must be scheduled. Slurm’s block/segment model lets TP/EP‑heavy jobs stay inside NVLink domains, trading wait time for throughput; aligning segment sizes with your parallelism plan can reduce tail latencies and requeues. Share the scheduling guidance with your infra owner. ²

For AI product leads, latency isn’t only about GPUs. Nvidia’s Vera CPU emphasis and agentic harness work show that tool calls, streaming events, and CPU‑bound steps shape UX and cost; verify your vendor supports streaming tool dispatch and reasoning segmentation so partial results arrive as they’re decoded. ¹

Finally, bake observability into training SLOs. NCCL Inspector’s Prometheus export gives you per‑collective, per‑rank visibility; the 13% throughput dip under network constraints is the kind of issue you can catch early with Grafana panels before it burns a week of experiments. ³

Action Items

1. Write a 2‑sentence brief on Nvidia’s $40B push: Summarize the OpenAI, Corning, and IREN stakes and share with your product and procurement leads to align on capacity and vendor strategy this quarter.
2. Share Slurm block scheduling guidance with infra: Ask your cluster admin to review --segment settings for your next TP/EP run so jobs stay inside NVLink domains where it matters.
3. Ask your AI vendor about streaming tool calls: Confirm support for streaming tool dispatch and reasoning segmentation (Anthropic‑compatible flows) to cut perceived latency in agent features.
4. Check your inference API rate limits: If you use Nvidia NIM or similar services, confirm current RPM quotas and request increases early to avoid throughput bottlenecks during launches.