Grok 4.7 Underperforms While OpenAI Agent Breaches Australian Medicare

Today we're going deep on a story that doesn't have a flashy model name attached to it, but might matter more than any of the launches this week: CoreWeave's argument that AI infrastructure has quietly become one giant computer, and that this shift is rewriting the rules underneath every agent you're using. **Technical Deep Dive** The core idea, straight from CoreWeave's Chen Goldberg, is this: a modern AI cluster is no longer a pile of GPUs you rent by the box. It's one enormous supercomputer, where compute, networking, storage, cooling, and software all have to behave as a single coordinated machine.

When you've got hundreds of GPUs working on the same training or inference job, a slow chip, a marginal network link, or a cooling hiccup doesn't just degrade one machine — it drags down the entire workload. That's a fundamentally different failure model than the "spin up a VM, it either works or it doesn't" cloud computing we've had for twenty years. The more interesting wrinkle is what agents do to this picture.

A chatbot request finishes in seconds. An agent can run for hours, make hundreds of tool calls, degrade quietly over time, and needs the infrastructure itself to observe, heal, and course-correct while it's still running. Goldberg notes that over 90% of CoreWeave's AI workloads still run on Kubernetes — so it's not dead — but the underlying resource model has to evolve to handle these tightly coupled, long-running, self-modifying jobs.

That's a very different orchestration problem than scheduling stateless web requests. **Financial Analysis** This matters financially because it reframes what you're actually paying for. CoreWeave and its rivals aren't selling GPU-hours anymore in any meaningful sense — they're selling guaranteed system-level throughput, which means the capex has to go into networking fabric, liquid cooling, and software orchestration just as much as into the chips themselves.

That's a much harder thing to price and a much harder thing to scale predictably. It also reframes cost conversations happening elsewhere in AI right now. Look at Shopify's story from Joanna's intel: a 96% cost reduction by moving off a frontier model onto something narrow and specialized.

That's the demand side of this equation — enterprises are learning that frontier intelligence is often massive overkill, and cheaper task-specific models can do the job for a fraction of the compute. Combine that with CoreWeave's supply-side reality, where infrastructure providers are racing to build systems capable of supporting messy, long-running agentic workloads, and you get a market where efficiency gains on one side are colliding with rising system complexity on the other. And zoom out further: AMD just crossed a trillion-dollar market cap largely by soaking up the overflow demand Nvidia couldn't fill.

The infrastructure buildout itself has become its own trillion-dollar economy, separate from whether any single model wins the benchmark wars. **Market Disruption** Competitively, this is a direct challenge to the old hyperscaler playbook. AWS, Azure, and Google Cloud built empires on the assumption that compute is fungible — rent a box, run your job, release the box.

CoreWeave's pitch, backed by its multi-rack Nvidia Vera Rubin NVL72 deployments, is that AI-native infrastructure requires purpose-built systems where the whole cluster is treated as one machine from day one. That's part of why neoclouds keep raising money and even IPO'ing — London's Nscale just filed for a New York listing targeting a $35 billion valuation, despite a $1 billion first-half loss. There's a security dimension here too, and it connects directly to two other stories in Joanna's feed today.

Claude Code just moved its cloud sessions into general availability on Anthropic-managed, persistent VMs — meaning agents now run with Anthropic's network permissions rather than your local sandbox, even after you close the window. And the unconfirmed Plugin4Shell vulnerability allegedly hitting Claude Code, Copilot, and Gemini CLI is exactly the kind of failure mode Goldberg is describing: once agents run persistently across shared infrastructure, a single weak link — in the network, the orchestration layer, or a plugin's SHA pinning — can compromise far more than one session. Infrastructure-as-single-supercomputer means security has to travel with the agent through the entire stack, not sit in one layer.

**Cultural & Social Impact** There's a quieter, more human thread in this story too. Goldberg jokes that if you like writing code, you may not be the one writing it much longer — one engineer on her team built something in three weeks that would have taken a year previously. That's not just a productivity stat, it's an identity shift for an entire profession.

What does "software engineer" mean when the code is largely generated, and your actual job is architecting the system, validating outputs, and managing the infrastructure the agent runs on? Goldberg frames the payoff as democratization: making advanced AI infrastructure accessible to people with different professions and locations, rather than keeping it concentrated among a handful of frontier labs. That's an optimistic read, but it cuts against a countervailing trend — the capital and engineering complexity required to build "one giant supercomputer" infrastructure is enormous, which could just as easily concentrate power among the few players who can afford to build it right, CoreWeave, Microsoft, Google, and Amazon among them.

**Executive Action Plan** So what should business leaders actually do with this? First, if you're running or evaluating agentic AI workloads, stop thinking about cost in dollars-per-token and start thinking in cost-per-completed-task — that includes retries, human rescues, and elapsed time, not just raw model spend. Shopify's 96% savings didn't come from a cheaper token rate alone; it came from matching the task to the right-sized model entirely.

Second, audit your assumptions about where your agents actually run. If you're using cloud-based coding agents like Claude Code, Copilot, or Gemini CLI, understand that persistent VM architectures mean your agent may be operating with the vendor's network permissions long after your session ends — and unconfirmed reports of vulnerabilities like Plugin4Shell mean you should be treating agent supply chains with the same scrutiny you'd give any production dependency. Third, don't assume your current cloud vendor's infrastructure was built for tightly coupled, long-running agentic jobs.

Ask specifically how they handle straggler mitigation, self-healing, and observability at the cluster level — because as Goldberg puts it, you're not renting parts anymore, you're operating one giant computer, and the weakest link in that machine is now your business risk.