Broadcom and Apollo Create $35B Platform for Deploying Anthropic AI Compute
Broadcom, Apollo, and Blackstone have formed a $35B financial platform to support the deployment of over 20 GW of AI compute capacity. Anthropic is the first customer, receiving specialized XPUs and networking capabilities.
Financial engineering, not technology: How Broadcom sold the future of AI for $35B in private debt
Insider analysis from an AI infrastructure finance observer
[The Gist]: What's Really Happening
On June 9, 2026, Broadcom, Apollo, and Blackstone announced the creation of the AI XPV (AI eXpansion Vehicle) platform with an initial tranche of $35B to deploy over 20 GW of compute capacity by 2028. Anthropic is the first customer, receiving over 1 GW of infrastructure based on Broadcom XPUs and networking solutions, with launch in mid-2026 at Fluidstack sites. It sounds like a technological breakthrough. In reality, it's the largest AI infrastructure securitization deal in history, shifting risks from Broadcom to creditors.
What's really happening here? Broadcom didn't sell chips. They sold a financial structure in which Apollo and Blackstone provide debt financing backed by Anthropic's future contracts. The scheme is simple but brilliant in its audacity: Broadcom gets money to produce XPUs now without diluting shareholders or touching its balance sheet. Apollo and Blackstone earn interest income. Anthropic gets compute capacity it would never have had under normal venture funding.
But there's a nuance that makes this deal historic. According to S&P Global Ratings, the first tranche is structured as two debt categories: $30B in senior notes A1/A2 with residual value support from Broadcom (i.e., Broadcom guarantees them with its assets) and $4.5B in a junior tranche B without such backing. S&P has already called this "credit-negative" for Broadcom but maintained its A- rating. This means the rating agency sees risk but believes Broadcom can weather it.
The essence: Broadcom has transformed from a chip supplier into a banker. They no longer just sell XPUs—they structure deals, raise capital, and distribute risks among Blackstone's insurance units and Apollo's pension funds. This is vertical integration not in technology, but in finance.
Timeline and Context
The deal didn't emerge from nowhere. It resulted from a two-year race for compute capacity, in which Anthropic, OpenAI, and other frontier labs realized their growth is limited not by algorithms but by access to electricity and chips.
| Date | Event | Amount / Details | Strategic Significance |
|---|---|---|---|
| 2024 | Anthropic signs 1 GW contract with Google | Using TPUs | Start of the "capacity race" |
| 2025 | Anthropic announces $50B partnership with Fluidstack | Multi-phase deployment | Alternative to Google dependency |
| April 2026 | Anthropic signs deal with Google and Broadcom for several GW of TPUs | ~3.5 GW from 2027 | Broadcom becomes strategic partner |
| June 9, 2026 | Launch of AI XPV Platform with $35B tranche | Apollo lead, Blackstone anchor | First securitization of AI infrastructure |
| June 2026 | S&P calls first tranche credit-negative | Broadcom A- rating maintained | Market risk recognition |
| Mid-2026 | Deployment of first GW for Anthropic at Fluidstack sites | Operational launch | Key test for the entire structure |
| 2027+ | Commissioning of capacity for OpenAI and other frontier labs | Up to 20 GW by 2028 | Platform scaling |
Key takeaway from the timeline: this deal is a hedge against Nvidia. Broadcom can't compete with Nvidia in the GPU training market (H100, B200). But they can offer an alternative for inference and specialized workloads via XPUs, and crucially, they can offer financing that Nvidia doesn't provide. Anthropic pays not for hardware but for "capacity as a service," and this capacity will be deployed at Fluidstack sites, which essentially become "contract factories" for AI labs.
Who Wins and Who Loses
Winner #1: Anthropic. This company just secured over 1 GW of compute capacity without an immediate cash outflow. Their projected annual revenue after these partnerships will reach $30B. They are preparing for an IPO, and having guaranteed compute capacity for years ahead is an asset that public investors will value. But there's a risk: if Anthropic can't monetize this capacity, the debt remains, and they are the primary borrower through the platform structure.
Winner #2: Apollo and Blackstone. They gained access to the largest private credit deal in AI infrastructure. Apollo leads the credit structure, Blackstone acts as anchor investor through its credit and insurance division. Their clients (pension funds, insurance companies) get returns tied to AI economy growth. If Anthropic and OpenAI indeed grow to $30-50B in revenue, the yield on these credit instruments will be significantly higher than traditional corporate bonds.
Winner #3: Fluidstack. This little-known AI cloud provider suddenly became the operator of one of the world's largest compute clusters. The first GW of Anthropic capacity will be deployed at their sites starting mid-2026. Fluidstack didn't just get a contract—it got validation at the level of Broadcom, Apollo, and Blackstone. The company's valuation will likely increase 3-5x after this announcement.
Loser #1: Nvidia (strategically). This deal directly bypasses Nvidia's monopoly on AI compute. Anthropic will use Broadcom XPUs (which include, among others, Google TPUs) and Broadcom networking solutions, not Nvidia GPUs. If the model proves successful, other frontier labs (OpenAI, Meta, xAI) will follow suit, and Nvidia will lose not only sales but also influence over standard-setting.
Loser #2: Broadcom bondholders (short-term). S&P called the first tranche credit-negative. This means Broadcom's debt burden increases, and the deal structure (especially the $4.5B junior tranche B without support) creates risk for creditors in case of Anthropic default. For now, the A- rating is maintained, but if the platform expands, rating agencies may revise it downward.
Loser #3: Small AI startups. This $35B tranche is for top players (Anthropic, OpenAI). Resources (electricity, chips, financing) are consolidating among leaders. Small startups that can't offer similar revenue levels will be left without access to compute capacity, widening the gap between the "golden billion" of AI labs and everyone else.
What the Media Isn't Saying
Insight #1: 20 GW is not marketing; it's an existential problem for the US power grid. 20 GW is roughly the capacity of 20 nuclear reactors. For comparison, Gartner forecasts that by 2026, the total capacity of all data centers worldwide will reach 132 GW. So one Broadcom platform aims for 15% of global capacity growth in the coming years. But where will all that electricity come from? Data centers already consume 565 TWh per year, and by 2030, Gartner predicts they will reach 290 GW capacity. Each new GW from Broadcom must be connected to grids somewhere, creating competition with residential areas, industry, and renewable energy. The platform hasn't solved this problem—it has simply shifted it to grid operators.
Insight #2: The $4.5B junior tranche B is a "toxic tail" that could be resold to unsuspecting investors. In the deal structure, $30B are senior notes with Broadcom support, and $4.5B is a junior tranche without support. In the securitization world, such "toxic tails" are often bought by investors who either don't understand the risk or are willing to speculate on it. If Anthropic can't service the debt, these $4.5B will be the first to lose value. The question: who bought them? Press releases don't disclose this, but I suspect they may have been sold through Blackstone structures to institutional investors with lower risk awareness.
Insight #3: Why is OpenAI the second client but not mentioned in details? The platform press release says it's intended for Anthropic and OpenAI, but all details (amounts, timelines, architecture) are given only for Anthropic. This means negotiations with OpenAI are not yet complete, or the deal structure for them will be different. OpenAI already has its own partnership with Broadcom to develop a custom AI chip, and they may not want to enter the same debt structure as Anthropic. This creates interesting dynamics: the platform is formally open to OpenAI, but in practice, Anthropic is the first and so far only beneficiary.
Insight #4 (non-obvious): Broadcom is hedging against a drop in XPU demand. Broadcom currently has $73B in AI order backlog. But what happens if demand falls? They would be left with factories and contracts. The AI XPV platform shifts this risk to creditors: if Anthropic doesn't need that much capacity, the debt remains, but the chips don't. Broadcom has already received money for the tranche. This is an ideal hedge: in case of growth, they share in profits through production; in case of decline, creditors bear the losses.
Forecast: Next 30 Days and 90 Days
Next 30 Days (by mid-July 2026)
Expect reactions from rating agencies Moody's and Fitch. S&P has already spoken, but Moody's and Fitch are silent so far. They will likely follow S&P's lead but without downgrading—just a comment about "risks associated with new financial instruments." If even one agency changes its outlook from "stable" to "negative," Broadcom's stock could fall 5-10%.
Also expect an announcement about Fluidstack's first project. Where exactly will the first GW of capacity be located? This is critical for understanding the platform's geography. If in the US, it's a win for the domestic market. If in Europe (e.g., amid the UK's sovereign AI plan), it signals global scale. If in Saudi Arabia or the UAE, it's a geopolitical choice.
Key indicator: publication of the junior tranche B terms. If they appear in public sources (e.g., prospectus), investors can assess real risk. If the information remains confidential, it means the structure is too risky for public disclosure.
90-Day Horizon (by September 2026)
By September 2026, we will likely see an announcement of the platform's second tranche. The first tranche of $35B is just the beginning. The platform targets 20 GW by 2028. If the first tranche provides 1 GW for Anthropic, then reaching 20 GW will require 19 more similar or larger deals. Apollo and Blackstone have already stated a "scalable structure for future deployments." The second tranche could be announced as early as September and will likely target OpenAI or possibly Meta (which is also working with Broadcom on its AI accelerator MTIA).
Also likely is an announcement of a partnership with a major energy giant (e.g., NextEra Energy or Shell). 20 GW is not just chips; it's electricity. Broadcom, Apollo, and Blackstone don't build power plants. They need a partner to provide capacity. Such an announcement could be the next step after the first tranche is launched.
Geopolitical scenario: China, seeing the platform's scale, may accelerate its own AI infrastructure financing programs through state banks. This would create asymmetry: in the West, private credit at high interest (Apollo, Blackstone); in China, state subsidies at low interest. If Chinese frontier labs (SenseTime, Baidu, Alibaba) gain access to cheaper financing, they could scale faster despite technological constraints. The US and Europe might respond by creating "sovereign AI funds" similar to the CHIPS Act but for infrastructure rather than manufacturing. In such a scenario, the AI XPV Platform could turn out not to be a "breakthrough" but a "prelude" to a full-scale public-private race for AI capacity.
Final assessment: Broadcom, Apollo, and Blackstone have not created a technology platform. They have created a financial mechanism that allows shifting the risks of AI expansion from tech company balance sheets to credit investor balance sheets. This is brilliant for Broadcom (they got $35B without dilution) and for Apollo (they got returns tied to AI growth). But it's risky for the entire ecosystem: if Anthropic doesn't meet revenue forecasts, the $4.5B junior tranche B will lose value, and investors will lose money. If Anthropic does meet them, this will become the new normal: AI labs no longer need venture capital; they need private debt backed by future capacity. We stand on the threshold of an era where "compute capacity" becomes as tradable an asset as oil or gas. And the first to understand this were not technologists, but financiers.
— Editorial Team
No comments yet.