In partnership with
Anthropic open-sourced Claude Code this week, accidentally.
A missing .npmignore entry shipped 512,000 lines of unobfuscated TypeScript to the npm registry. Within hours, the code was mirrored, dissected, rewritten in Python, then rewritten again in Rust. The clean-room rewrite hit 100,000 GitHub stars in a day, the fastest-growing repository in the platform's history. Days earlier, Anthropic's unreleased "Mythos" model also turned up in a public data cache.
Two leaks in one week.
Not a great look for the lab that built its brand on being careful.
I'm Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
The GPU Audio Companion Issue #100
Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.
Today's issue is brought to you by Rafay
The GPU lives on sponsorships. If you value independent, no-fluff analysis of the AI infrastructure market, the sponsors are who make it possible. I'd appreciate you checking them out.
How to Build an AI Factory with Rafay
Organisations are investing heavily in GPUs, but infrastructure alone doesn’t create an AI factory. That’s why Rafay put together this paper exploring the architectural and operational model required to transform raw GPU infrastructure into a self-service AI platform. Click the button below to learn how leading companies operationalise AI infrastructure with governance, developer self-service, and production-ready AI workflows.
Anthropic Leaks Claude Source Code and Next-Generation Model in One Week
What do you get if you combine 512,000 lines across 1,900 files with one missing config entry? The biggest accidental code disclosure in AI history.
Anthropic shipped a 59.8MB source map inside Claude Code's npm package on March 31, exposing the full unobfuscated TypeScript source of its flagship coding agent. No hack. No reverse engineering. A build configuration error. An X post from security researcher Chaofan Shou hit 22 million views within hours.
Anthropic called it "human error, not a security breach", scrambling to issue DMCA takedown requests.
The developer community moved faster than Anthropic's lawyers.
A clean-room Python rewrite called "claw-code" hit 50,000 GitHub stars in two hours and crossed 100,000 in a single day, the fastest-growing repository in the platform's history. It was subsequently rewritten in Rust with xAI providing Grok credits to assist with the port. Further separate forks then stripped telemetry, unlocked experimental features, enabled use with competing models, mirroring versions of the codebase across decentralised platforms beyond Anthropic’s legal reach.
The leak also revealed autonomous "dreaming" agents, DRM-like client attestation, a multi-agent framework, internal benchmarks, model codenames, and anti-competitive countermeasures against rival coding tools.
Why this matters:
Competitors just got the full production source code of Anthropic’s $2.5 billion in revenue crown jewel, complete with prompt engineering, agent orchestration, tool wiring, and the internal benchmarks Anthropic uses to measure itself against OpenAI's Codex and Google's Gemini CLI. Both of which, incidentally, are already open source.
The clean-room rewrite creates a legal template that could reshape code IP in the AI era. A Korean developer rewrote Claude Code in Python using AI tools before sunrise. If AI-assisted rewrites constitute new creative works outside DMCA reach, every proprietary coding tool is one leaked source map away from forced evolution into an open-source competitor.
Anthropic's brand is safety, caution, and responsible scaling, but just days before this leak, Fortune revealed "Mythos," Anthropic's next model described internally as a "step change," in a publicly accessible data cache. Two accidental leaks in seven days, the second caused by the very toolchain (Bun) it acquired in late 2025, actively undermine that brand. The commercial impact may be limited, but the question of whether operational security matches the rhetoric is no longer theoretical.
CoreWeave Closes First Investment-Grade GPU Debt. OpenAI Closes Largest Private Fundraise Ever.
$130 billion in AI capital deployed in one week. One deal proved GPUs are bankable. The other proved circularity is a feature, not a bug.
CoreWeave closed an $8.5 billion delayed draw term loan rated A3 by Moody's. The sum, while not huge, is the first investment-grade GPU-backed financing in history. Bloomberg reports it's backed by Meta contracts worth $19 billion+. Blackstone anchored. SOFR + 2.25% floating, 5.9% fixed. Total CoreWeave financing: $28 billion in 12 months. CEO and CDO both sold shares the same week.
Separately, OpenAI closed $122 billion at an $852 billion valuation. Amazon committed $50 billion (now the exclusive cloud partner). NVIDIA committed $30 billion. SoftBank committed $30 billion. Revenue: $2 billion per month.
Why this matters:
Two years ago, GPU-backed lending was Sandton charging Arkon 17.5% with steep discounts (Issue #99). Now CoreWeave borrows at SOFR + 225bps with an A3 rating. If this structure is repeatable, investment-grade debt markets open to any neocloud with a hyperscaler anchor tenant.
OpenAI's investor list is its vendor manifest. Amazon's $50 billion buys exclusive cloud. NVIDIA's $30 billion flows back through GPU purchases. SoftBank's $30 billion funds Stargate. The capital is circular by design. $852 billion at 35x revenue, with $14 billion in projected 2026 losses.
$130 billion in AI capital commitments in a single week. One deal prices GPU infrastructure as institutional fixed income. The other prices a single AI company above the GDP of most countries.
A company with $1.57 million in annual revenue just signed a billion-dollar GPU cloud contract.
Sharon AI disclosed Canva as its first major customer this week, alongside a $1.25 billion five-year deal with ESDS Software Solutions for 8,200 NVIDIA B300 GPUs. Shares surged 24%. Sharon AI reported $1.57 million in 2025 revenue and a $39.8 million net loss. It listed on Nasdaq in February with 432 GPUs live (Issue #91). Six weeks later: Canva won, $1.25 billion contracted, 70MW committed with NEXTDC, and Australia's first Cisco Secure AI Factory operational with NVIDIA (Issue #93).
Why this matters:
Canva is one of Australia's most valuable tech companies. Landing it as a lighthouse GPU cloud customer within weeks of IPO, from a starting fleet of 432 GPUs, makes the APAC sovereign AI thesis tangible rather than theoretical. The demand exists. Someone local is capturing it.
$1.25 billion contracted against $1.57 million in revenue. Sharon AI is selling capacity it hasn't deployed, on hardware it hasn't procured, in facilities it doesn't yet occupy at scale. That's the neocloud model in its purest form: contracted demand first, then capital, then hardware. The execution risk is obvious. The commercial signal is harder to dismiss.
Australia now has two NVIDIA Cloud Partners: Sharon AI (70MW, enterprise customers including Canva) and Firmus ($10 billion in Blackstone financing, 1.6GW target, hyperscaler contracts). Different scale, different customer segment, same thesis. APAC sovereign compute is no longer a single-company bet.
NVIDIA Invests $2 Billion in Marvell to Pull Custom ASICs into NVLink Fusion
NVIDIA just invested in the company that designs chips for its biggest customers' alternatives.
NVIDIA has invested $2 billion in Marvell Technology and has signed a partnership that connects Marvell's custom XPUs to NVLink Fusion. Marvell designs ASICs for AWS (Trainium), Microsoft, and Google. Under the deal, Marvell provides custom silicon and networking; NVIDIA supplies Vera CPUs, ConnectX NICs, Bluefield DPUs, and Spectrum-X switches. Every NVLink Fusion platform requires at least one NVIDIA component. AMD, Intel, and Broadcom remain absent, backing the open UALink standard.
Why this matters:
Marvell's entire custom silicon business exists to reduce NVIDIA dependency. By bringing Marvell into NVLink Fusion, NVIDIA ensures those ASICs continue to generate NVIDIA revenue per rack. The competitor becomes a channel partner.
The interconnect war is splitting the industry. NVLink Fusion (Marvell, Samsung Foundry, Arm) on one side, UALink (AMD, Intel, Broadcom) on the other. One has shipped. The other hasn't. The camp you join determines your infrastructure economics for a decade.
$2 billion into Marvell follows $2 billion into CoreWeave, $2 billion into Nebius, $2 billion into Synopsys. The pattern across all four: capital deployed as ecosystem lock-in.
Google Releases Gemma 4 Under Apache 2.0: Four Sizes, Runs on a Phone
Google just gave away a model family that outcompetes offerings 20x its size.
Google released Gemma 4 under the Apache 2.0 licence. Four sizes: 31B dense (#3 on Arena), 26B mixture-of-experts, and two edge models (E4B, E2B) that run on phones, Raspberry Pi, and Jetson Nano with near-zero latency. Native multimodal (video, image, audio), 140+ languages, function calling built in, 256K context window on the larger models. NVIDIA is already optimising Gemma 4 for RTX. Built from the same research as Gemini 3.
Why this matters:
Apache 2.0 means any developer, company, or government can deploy Gemma 4 commercially without licensing fees. In turn, Google commoditises the model layer while monetising the infrastructure underneath: TPUs, Cloud, and the Android ecosystem that Gemma 4 will power via Gemini Nano.
A 31B model ranking #3 on Arena, available for free, resets the floor for any enterprise evaluating model costs. For production workloads that don't need frontier capability, the commercial API business just got harder to defend.
The edge models are the strategic play. E2B runs on a phone. Google worked with Qualcomm and MediaTek on hardware optimisation. When the most capable open model runs locally, inference revenue shifts from cloud API calls to device silicon. Google doesn't need to win the cloud inference war if it wins the on-device one.
Oracle Cuts 30,000 Jobs to Fund AI Data Centre Expansion
The human cost of the capex cycle keeps rising.
Oracle is cutting up to 30,000 jobs to redirect capital into AI infrastructure. The company took on $58 billion in new debt in two months, including a $50 billion bond offering in February. It has committed to 4.5GW of Stargate capacity for OpenAI, a partnership exceeding $300 billion over five years. Larry Ellison told investors AI models are now writing "a lot of the code that Oracle is writing." Restructuring costs: $2.1 billion.
Why this matters:
Oracle stood next to Trump to announce 100,000 jobs. Those are construction workers. Temporary. The 30,000 being cut are software engineers, cloud architects, and SaaS operators. When infrastructure becomes the primary asset, labour gets repriced.
$58 billion in new debt. Banks are pulling back. Oil above $100. Borrowing costs are rising amid the Iran war. Oracle is cutting headcount to free up $8-10 billion in cash flow, as capital markets are tightening at the exact moment capex commitments are accelerating.
Oracle is reportedly exploring selling Cerner (purchased for $28.3 billion in 2022) and asking new customers to bring their own chips. Everything that isn't a data centre is being sold, cut, or deprioritised.
Starcloud Raises $170 Million at $1.1 Billion for Orbital Data Centre Constellation
Space-based AI compute hit unicorn status this week. The satellite count in the press release is 88,000. The number currently in orbit is one.
Starcloud raised $170 million at $1.1 billion, led by Benchmark and EQT Ventures. The company plans an 88,000-satellite data centre constellation and is working with NVIDIA, AWS, and Google Cloud. It launched an H100 into orbit in November 2025. A second launch carrying AWS Outposts is planned for October. Earlier investors include In-Q-Tel, the CIA's venture arm. Cost parity with terrestrial facilities expected by 2028-29.
Why this matters:
SpaceX/xAI filed for one million data centre satellites. Blue Origin filed for orbital data centres. Starcloud is the first independent startup to hit unicorn on the same thesis. Three companies now bet terrestrial power and land constraints make orbit viable within five years.
The H100 in orbit is the proof point every other announcement lacks. Starcloud has run AI training and inference on NVIDIA hardware in space. The October AWS Outposts launch would put a hyperscaler's edge stack in orbit for the first time.
Cost parity by 2028-29 depends on Starship reaching price points SpaceX promises but hasn't delivered at volume. Plausible thesis. Aggressive timeline. One operational satellite.
The Rundown
Demand is the story. Everything else is a symptom.
CoreWeave got an A3 rating because Meta needs $19 billion of GPU capacity, and Moody's believes the cash flows. OpenAI raised $122 billion because Amazon, NVIDIA, and SoftBank all need what it's building and want to own a piece of the supply chain that serves it. Oracle fired 30,000 people because the data centre contracts are worth more than the engineers. Sharon AI signed a billion-dollar contract because Canva needed GPUs in Australia and couldn't get them from anyone else. A company with one satellite hit a billion-dollar valuation because the terrestrial grid is full.
And Anthropic accidentally gave the entire developer ecosystem a blueprint for building coding agents.
That last one may yet matter the most.
OpenClaw arguably kicked off this demand wave. And that was as a security nightmare. Claude Code is a different beast. That beast has been laid bare. Its architecture was rewritten in Python and Rust within hours, starred 100,000 times in a day, and will likely accelerate the open-source agent ecosystem faster than any deliberate release could have. Google shipping Gemma 4 under Apache 2.0 the same week, only compounds it. The tools to build AI coding agents are now free, documented, and multiplying.
A hundred issues in, and demand shows no signs of slowing down.
See you next week.