LLM Hosting Roundup: Anthropic’s capacity crunch, Vera Rubin and DeepSeek’s $15B (23 Oct 2026)

This week's coverage points to a hosting market under pressure from two directions: frontier providers are securing enormous amounts of compute, while infrastructure vendors are redesigning racks, cooling, power delivery and serving software for agentic workloads. The clearest operational theme is that throughput, capacity and energy efficiency are becoming inseparable for teams evaluating hosted inference.

• Anthropic's cloud capacity challenge: TechCrunch reported that Anthropic is leasing its own data centers and negotiating cloud contracts while seeking enough capacity from AWS, Google and Microsoft.
• NVIDIA's factory-built AI infrastructure: ServeTheHome covered the DSX POD reference design, which packages seven Vera Rubin NVL72 racks with the supporting systems needed for deployment.
• Dell adds agentic-inference systems: Data Centre Dynamics reported Dell's Vera CPU systems and its first Groq 3 LPU rack-scale inference server.
• Inference efficiency benchmarked on Vera Rubin: Artificial Analysis reported up to roughly 5.7× the throughput per megawatt of GB300 NVL72 on one DeepSeek workload.
• DeepSeek reportedly raises $15B: TechCrunch said the funding would support increased spending on AI infrastructure and open models, while noting that the lab did not comment.
• Another reported Anthropic–Nscale compute deal: TechCrunch reported that talks for a further $45B deal were underway and could change.
• Reflection launches Beam: Artificial Analysis assessed the open-weights model at 501B total and 23B active parameters, with a 1M-token context window.
• Three physical infrastructure constraints: In partner content supplied by xFusion, The Next Platform discussed cooling, power delivery and interconnect as coupled constraints for dense AI systems.

Anthropic's cloud-capacity challenge

TechCrunch reported, based on a leaked S-1 filing cited by The Wall Street Journal, that Anthropic is having difficulty securing enough cloud capacity from AWS, Google and Microsoft as those companies increasingly view the Claude maker as a competitor as well as a customer. The report said Anthropic is leasing its own data centers and working with third-party operators, and that the filing described potential cloud-contract spending of more than $750 billion, including deals still under negotiation. TechCrunch also reported $16.3 billion in first-half revenue and an $80.7 billion net loss, while noting that Anthropic did not immediately respond to its request for comment. [1]

For hosting buyers, the evidence highlights the scale and complexity of capacity planning for frontier services: supply may involve hyperscalers, owned facilities and additional operators rather than a single cloud relationship.

NVIDIA's DSX POD turns AI-factory deployment into a packaged system

ServeTheHome reported that NVIDIA's DSX POD reference design places seven Vera Rubin NVL72 racks, plus networking, cooling and management hardware, inside a factory-assembled shipping container. The article said one POD could provide 25 EFLOPs of AI inference, while also qualifying that the container still needs external power, networking and a larger cooling system. NVIDIA expects partners including Gigabyte, HPE, Dell, Delta and QCT to begin delivering DSX PODs worldwide in early 2027; NVIDIA is providing the reference design rather than building or selling the PODs itself. [2]

The same coverage described NVIDIA's OpenShell reference architecture for agentic safety: Vera Rubin handles LLM reasoning, Vera CPUs handle agent execution and the OpenShell runtime, and BlueField-4 DPUs provide hardware security enforcement. The deployment model is relevant to hosting teams because it combines inference compute, orchestration and security into a repeatable infrastructure design rather than treating them as separate installations.

Dell adds Vera CPU systems and Groq 3 rack-scale inference

Data Centre Dynamics reported that Dell expanded its AI Factory portfolio with the PowerEdge M9882, a 48U rack containing 256 NVIDIA Vera CPUs; the air-cooled, two-socket R8822 server; and the PowerEdge XE9810, Dell's first system featuring Groq 3 LPUs. Dell said the M9882 targets agentic workloads such as orchestration, tool use, code execution and evaluation alongside GPU infrastructure, while the XE9810 is intended as a disaggregated complement to GPU clusters for production agentic inference. [3]

The R8822 was reported as available in October, while the M9882 and XE9810 were scheduled for Q1 2027. The article did not provide the number of LPUs per XE9810 tray, so that configuration remains unspecified.

Vera Rubin shows large gains on an agentic-inference benchmark

Artificial Analysis reported initial AA-AgentPerf results in which NVIDIA's Vera Rubin NVL72 delivered up to roughly 5.7× the total throughput per megawatt of GB300 NVL72 on DeepSeek V4 Pro, with reported gains of roughly 5.4× on DeepSeek V4.1 Flash, 3.0× on GLM-5.2 and 2.8× on Kimi K3 at selected per-user speeds. The system tested had 72 Rubin GPUs and 36 Vera CPUs. [4]

The benchmark is designed around coding-agent trajectories rather than fixed chat prompts: requests can carry up to 132K input tokens, 97% of input tokens were repeated from the prior request on average, and tool-call pauses were included. Artificial Analysis said the results used techniques including KV-cache reuse, session-aware routing with NVIDIA Dynamo and disaggregated serving. It also qualified that the reported power measurements covered GPU power under load and excluded CPUs, networking and cooling, so the figures are not a full-facility operating-cost comparison.

DeepSeek reportedly raises $15B for infrastructure and open models

TechCrunch reported, citing Bloomberg, that DeepSeek raised about $15 billion—roughly 100 billion yuan—two days after U.S. agencies were ordered to remove the lab's AI technology from their systems over security concerns. The report said potential investors included CATL and Tencent, and that DeepSeek was ramping up spending on AI infrastructure and open models for domestic and international markets. DeepSeek did not respond to TechCrunch's request for comment. [5]

For the hosting market, the reported use of funds matters because it links model competition with additional infrastructure investment, while the reported restrictions show that deployment strategy can be shaped by policy as well as technical demand.

Anthropic and Nscale reportedly discuss another $45B compute deal

TechCrunch reported that Anthropic and AI infrastructure provider Nscale were said to be nearing another $45 billion compute deal, citing Bloomberg and people familiar with the matter. The sources cautioned that talks were still underway and details could change; TechCrunch said neither company had confirmed the report at publication. [6]

The report places the proposed agreement in the context of a broader series of large AI compute partnerships, but it does not establish that this proposed deal was finalized. That distinction matters when comparing announced capacity with capacity that is actually contracted and deployable.

Reflection launches the open-weights Beam model

Artificial Analysis reported that Reflection's first model, Beam, is an open-weights model with 501 billion total parameters and 23 billion active parameters, a 1-million-token context window and text-only modality. Its evaluation gave Beam a score of 30 on the Artificial Analysis Intelligence Index and 48.5% on AutomationBench-AA, while also reporting 3.0% on Terminal-Bench 4.0 and a 47% hallucination rate on AA-Omniscience. [7]

These are Artificial Analysis's evaluation results, not a provider-reported hosting benchmark. For teams assessing deployment options, the published parameter mix, context window and token-efficiency discussion provide useful model characteristics, while the mixed coding and hallucination results argue against treating one aggregate score as a complete serving decision.

Dense AI infrastructure faces coupled cooling, power and interconnect constraints

In partner content supplied by xFusion, The Next Platform described cooling, power delivery and interconnect as coupled constraints for dense AI infrastructure. The piece attributed a trajectory of 240 kW current-generation racks and a roadmap toward megawatt-class racks to xFusion's presentation at the 2026 OCP Global Summit, and discussed direct liquid cooling, 800V DC distribution and optical interconnect as responses. [8]

Because this is explicitly partner content supplied by xFusion, its figures and product claims should be read as the company's presentation rather than independent measurements. The operational point is still relevant to hosting: a rack can be limited by heat removal, power-delivery losses or network bandwidth, so solving only one bottleneck can leave compute stranded.

Conclusion

Two evidence-supported points stand out. First, frontier hosting is becoming a capacity-and-capital problem as providers pursue very large cloud and compute commitments while also building or packaging their own infrastructure. Second, serving economics increasingly depend on workload-aware systems—especially KV-cache reuse, routing, disaggregated serving, cooling and power design—rather than on accelerator specifications alone.

News Sources

1. TechCrunch - Anthropic's cloud partnerships are getting messy
2. ServeTheHome - NVIDIA GTC Berlin 2026 NVIDIA Intros DSX PODs and More
3. Data Centre Dynamics - Dell unveils Nvidia Vera CPU offerings, launches its first Groq 3 LPU rack-scale inference solution
4. Artificial Analysis - First AA-AgentPerf results for NVIDIA Vera Rubin NVL72
5. TechCrunch - DeepSeek reportedly raises $15B days after US restrictions
6. TechCrunch - Nscale and Anthropic are reportedly in talks for yet another $45B compute deal
7. Artificial Analysis - Reflection has released Beam, their first model launch
8. The Next Platform - Three Physical Walls Are Hitting AI Datacenter Infrastructure at the Same Time - And the Compound Effects Are Worse