xAI’s Grok 4.5 Trained on NVIDIA’s Most Powerful Rack-Scale AI System

NVIDIA’s corporate account on X posted congratulations to xAI on the launch, confirming the hardware behind the model and describing Grok 4.5 as “the first model trained specifically for coding and agents.” The post also referenced Cursor, a developer-focused coding environment, as part of the setup — though the exact nature of that integration hasn’t been fully detailed in official technical documents.

The announcement puts two of the technology industry’s biggest names in the same frame: NVIDIA as the infrastructure provider, and xAI — the AI company associated with Elon Musk, distinct from SpaceX the rocket firm — as the model developer.

What Is the GB300 NVL72?

The GB300 NVL72 is not a server you’d find in a typical office. It’s a rack-scale system — an entire cabinet — packing 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs into a single, fully liquid-cooled unit. It draws around 120 kilowatts of power per rack, which demands the kind of infrastructure only large, purpose-built data centres can provide.

NVIDIA says the system can deliver up to 50 times the overall AI factory output performance of its older Hopper-based platforms. Third-party technical analyses put the raw compute at around 1.1 exaFLOPS of dense FP4 performance — that’s more than one million trillion floating-point operations per second. All 72 GPUs are linked via fifth-generation NVLink, with aggregate bandwidth of around 130 terabytes per second, letting them behave as a single unified compute unit rather than separate cards.

The GB300 NVL72 succeeds the GB200 NVL72, offering roughly 1.5 times higher dense FP4 performance, twice-as-fast attention operations, and around 50 per cent more HBM3e memory per rack. Reported memory figures range between 20 and 40 terabytes per rack depending on configuration — the precise standard figure varies across technical sources and hasn’t been pinned down in a single official specification.

Why Coding and Agents?

Grok 4.5’s focus on software development and agentic tasks reflects a clear shift in how AI companies are positioning their models. Rather than general-purpose assistants, the push now is towards systems that can write and review code, interact with tools and APIs, and carry out multi-step tasks without constant human input.

That’s a different kind of workload from generating text. It demands fast, accurate reasoning over long sequences of instructions — exactly what NVIDIA says the GB300 NVL72 was built for. The system is designed for real-time inference on trillion-parameter models, the scale at which frontier AI systems now operate.

Software developers and IT professionals stand to gain new tools from models like Grok 4.5. But the same capabilities raise questions. Security researchers have flagged that highly capable coding models could be misused for malicious code generation or automated cyber attacks if governance isn’t tight. Labour market analysts have noted that agentic coding tools may shift demand in software development teams, placing a premium on workers who can direct and quality-check AI output rather than write every line themselves.

Concentration of Power

Critics have pointed to a broader pattern behind announcements like this one. A small number of companies — NVIDIA foremost among them — now supply the hardware underpinning nearly all frontier AI development. That concentration raises questions about market dominance and democratic oversight of technology that is increasingly shaping public services, employment and daily life.

Environmental concerns also follow large-scale AI infrastructure. A rack drawing 120 kW around the clock, replicated across a data centre, carries a big energy and cooling footprint. Planning and environmental bodies in the UK and elsewhere are paying closer attention to the resource demands of AI workloads as the industry scales.

Jensen Huang, NVIDIA’s chief executive, has described the GB300 NVL72 and its siblings as the foundation of what he calls “AI factories” — facilities that don’t just store data but actively produce AI output at industrial scale. That framing captures the ambition, but it also captures the resource intensity.

Access for Smaller Organisations

Most businesses won’t run a GB300 NVL72. The hardware cost alone would run to millions of pounds, before factoring in the power supply and liquid cooling infrastructure needed to keep it running.

But that doesn’t mean Grok 4.5 stays out of reach. XAI and NVIDIA are positioning the model for access via cloud platforms and software products — meaning a small firm or individual developer could use Grok 4.5’s capabilities through an API or a coding tool without ever knowing what hardware is behind it. That’s how most AI capability reaches the market: not through local hardware, but through services built on top of it.

Start-ups and SMEs could, in theory, access advanced agentic coding tools at relatively low cost through cloud APIs. The barrier isn’t the model — it’s knowing how to use it well.

What This Means for Kent Residents

No GB300 NVL72 installations in Kent have been confirmed, and the system’s power and cooling demands make local deployment unlikely for any but the largest facilities. Kent-based businesses, councils and residents are far more likely to encounter Grok 4.5 through cloud-based tools and software products than through any local data centre. For the county’s tech sector and its students studying computer science or digital skills, the more immediate question is how quickly agentic coding models like this one change what employers expect — and whether local colleges and universities move fast enough to keep curricula in step.