Broadcom’s Latest AI Bet: A $100 Billion Debt Pile

Chips & GPUs

Broadcom is in talks to raise over 60 billion dollars in debt, potentially reaching 100 billion, to finance AI chip capacity for Anthropic and other customers, with Blackstone and Apollo Global Management expected to participate.

AP

Staff

August 23, 2026 · 4 min read

Broadcom is in talks with a group of lenders to raise more than 60 billion dollars in debt for an AI chip financing deal that would directly benefit Anthropic and other customers, with the total package potentially reaching as much as 100 billion dollars once every tranche is counted. Private equity giants Blackstone and Apollo Global Management are reportedly in discussions to participate, extending a financing partnership the three firms first struck in June.

How the financing is actually structured

The proposed deal splits into two layers. A senior-secured tranche, ranked first for repayment and backed by specific assets, could run anywhere from 60 to 70 billion dollars, with Broadcom itself guaranteeing part of it. A junior tranche of roughly 30 billion dollars would sit beneath that. The debt would be issued through a special-purpose vehicle rather than sitting directly on Broadcom’s own balance sheet, and Broadcom’s guarantee on the senior portion is the mechanism that would let those top-tier notes secure investment-grade credit ratings, which meaningfully lowers the cost of borrowing at this scale. Talks remain ongoing, terms could still change, and the financing may ultimately roll out in stages rather than as one single transaction.

The structure closely mirrors an earlier, smaller version of the same arrangement: in June, Broadcom, Blackstone, and Apollo struck a 35 billion dollar deal specifically to expand Anthropic’s computing capacity, part of a broader partnership aimed at delivering more than 20 gigawatts of AI compute by 2028. This new financing effectively scales that same playbook up by roughly three times.

Why an AI lab needs a chipmaker to arrange its financing

The arrangement reflects a genuinely different financing model than a company simply buying chips outright. Anthropic and other AI labs need enormous, predictable compute capacity locked in well ahead of when they’ll actually use it, but committing that much capital directly strains even a well-funded lab’s balance sheet. By having Broadcom, backed by Blackstone and Apollo’s private capital, arrange and partially guarantee the debt, Anthropic can secure the compute commitment without carrying the full financing burden itself, while Broadcom locks in a customer for chips and networking equipment worth tens of billions of dollars over the life of the arrangement. Both sides get something specific: Anthropic gets guaranteed capacity, Broadcom gets guaranteed demand, and the private credit firms get a large, structured lending opportunity in a market segment growing faster than most others available to them right now.

That dynamic, private capital firms stepping in specifically because traditional banks are reaching the limits of what they’ll underwrite alone for AI infrastructure, is becoming a defining feature of how this specific buildout gets financed. Broadcom’s own visibility into future demand backs up the scale here: the company reportedly has line of sight into roughly 10 gigawatts of AI-related demand for 2027 from clients including Anthropic and Meta, a demand curve worth more than 8 million U.S. households’ worth of power consumption.

A competitive play against Nvidia, not just a financing story

Broadcom’s core business here is designing custom AI accelerators, ASICs, as an alternative to Nvidia’s general-purpose GPUs, the same category covered in our explainer on TPUs versus GPUs. Every dollar of financing that makes it easier for Anthropic to commit to Broadcom-designed chips is also, directly, a dollar of competitive pressure on Nvidia’s own AI chip dominance. That competition isn’t uncontested even within the ASIC space: Broadcom shares fell as much as 5.9 percent the day before this financing news broke, after rival chipmaker Marvell Technology disclosed its own new custom chip agreement with Alphabet covering AI inference accelerators, a reminder that the custom-silicon market Broadcom is trying to lock in demand for has genuine, well-funded competitors of its own.

What a deal this size actually signals

A financing package approaching 100 billion dollars would rank among the largest corporate debt raises tied to the AI buildout to date, and the scale itself is a signal worth reading carefully: it implies Broadcom and its lending partners expect AI compute demand, at least from customers like Anthropic, to keep growing steeply enough to justify locking in financing at this magnitude years in advance. Broadcom shares rose as much as 1.1 percent in after-hours trading once the talks were first reported, suggesting investors are reading the news as a sign of continued, durable demand rather than as a worrying sign of overextension, though the stock’s separate, sharper decline on the Marvell news the day before is a reminder that the custom-chip market’s growth is not being read as guaranteed to flow toward any single company.

See Bloomberg’s original reporting for further detail on the deal terms.

GitHub’s Outage Exposed How Dependent Coding Has Become on AI

Coding Assistants

GitHub's nearly eight-hour outage on August 17 revealed that Copilot runs on a genuinely separate authentication and routing layer than the rest of the platform, recovering hours after every other core service was already fixed.

AP

Staff

August 23, 2026 · 4 min read

GitHub went down for nearly eight hours on August 17, taking Copilot, Actions, Issues, Pull Requests, and enterprise authentication with it. In its own post-mortem, the company traced the root cause to network saturation on load balancers in its Central US data center, triggered by a new peak in traffic, not a bad deployment or configuration change. The outage itself is a familiar kind of incident. What makes it worth real attention is how it recovered: Copilot came back nearly three hours after every other core service was already fixed, exposing that the AI layer now runs on infrastructure with a fundamentally different failure and recovery pattern than the rest of the platform.

How the outage actually unfolded

The incident began at 13:28 UTC and ran until 21:15 UTC, seven hours and forty-seven minutes end to end. At its peak, web and API error rates hit roughly 20 percent, while archive downloads and raw repository content, the kind of requests triggered constantly by installation scripts, Docker builds, and CI pipelines pulling source files, saw error rates near 50 percent. SAML and OIDC authentication, SCIM provisioning, and Team Sync went down alongside Actions, effectively locking out any organization relying on single sign-on for its CI/CD pipeline, not just individual developers trying to push code.

Seven of the eight affected services, API Requests, Actions, Git Operations, Issues, Pages, Pull Requests, and Webhooks, were declared mitigated by 16:59 UTC, roughly three and a half hours in. Copilot was conspicuously absent from that recovery announcement. It did not fully clear until the incident closed entirely at 21:15 UTC, nearly two and a half hours after everything else had already stabilized. According to GitHub’s own status page, the specific reason traces to a separate technical failure: delayed replies to a single internal endpoint triggered a latent retry bug in VS Code that amplified Copilot-related traffic by roughly 10 times, generating additional load precisely while the system was already trying to recover.

Why Copilot lives on a genuinely different clock

The core technical detail worth understanding is that Copilot’s underlying AI model providers stayed operational throughout the entire incident. Whatever kept Copilot down after everything else recovered was not a failure upstream at the model layer, it sat entirely inside GitHub’s own authorization and routing infrastructure, the plumbing that connects a developer’s request to the model actually generating a response. That distinction matters for anyone building products on top of AI coding assistants: the model being available is necessary but not sufficient. The authentication, routing, and retry logic wrapped around it constitutes its own independent point of failure, one that can behave completely differently from the rest of a platform’s infrastructure even when they’re nominally part of the same product.

The capacity problem GitHub has been racing to solve

This outage did not happen in isolation. GitHub CTO Vladimir Fedorov wrote in April that the company set out last fall to expand capacity tenfold, then concluded by February it actually needed to plan for 30 times its current scale, a revision that reflects just how fast AI-driven coding activity has grown on the platform. GitHub’s own July infrastructure report noted that agentic development workflows have surged since late December 2025, with pull request activity now generating load that spans Git storage, Actions, search, and background job processing simultaneously, a fundamentally different traffic pattern than the platform was originally architected around. In June, a Microsoft spokesperson acknowledged that the spike in AI agent development had tested GitHub’s infrastructure limits directly, and said the company was accelerating its migration onto Azure while separately pursuing a multi-cloud strategy that includes renting capacity from AWS, Microsoft’s own biggest cloud competitor.

That last detail is worth sitting with: Microsoft’s own developer platform apparently needs capacity from a rival cloud provider to keep pace with AI-driven demand, a genuinely unusual admission for a company that would ordinarily prefer to keep that workload entirely on its own infrastructure. It is also GitHub’s 13th recorded incident in the first 17 days of August alone, according to independent tracking, a frequency that points toward a systemic capacity strain rather than an isolated, one-off event.

What this means for teams depending on AI coding tools daily

For any team that has made Copilot or a similar AI coding assistant a genuine part of its daily workflow, this outage is a concrete argument for having a fallback plan that does not assume the assistant will always be there, not because the underlying model is unreliable, but because the infrastructure wrapped around it is a separate, independently failing system with its own recovery timeline. GitHub says it is now working on isolating critical services from each other and designing systems that preserve partial functionality during future incidents, exactly the kind of architectural response this outage’s specific failure pattern, one service’s retry bug delaying a completely different service’s recovery, points toward as necessary.

See GitHub’s own incident status page for the complete technical post-mortem.

Loading more stories…

You've reached the end — that's every story we've published.