Broadcom, Apollo & Blackstone Launch $35B AI Compute Financing Platform for Anthropic

Asset-Backed AI Compute Arrives

Broadcom announced an AI XPV financing platform with Apollo and Blackstone that treats GPU infrastructure as a financeable asset class, launching with a $35 billion transaction to support Anthropic's expansion past 1 GW of compute capacity. The platform targets more than 20 GW of AI compute deployment by 2028, effectively transforming how enterprises will access and pay for inference and training capacity.

The initial deal covers Anthropic's data center deployment starting mid-2026 at sites provided by Fluidstack. Within days, KKR launched a parallel vehicle—Helix Digital Infrastructure—backed by more than $10 billion in commitments from Kuwait Investment Authority, NVIDIA, and Vistra to build AI data centers, power infrastructure, and network connectivity. Together, these moves redirect roughly $45 billion of private capital toward GPU infrastructure in direct competition with hyperscaler-owned capacity.

What Changed for Enterprise Buyers

The shift to long-term, asset-backed financing ends the illusion that AI compute will behave like elastic cloud resources. Capacity financed at this scale flows to anchor tenants—frontier labs and large clouds—under multi-year contracts. Enterprises buying inference or training capacity through Anthropic's API, AWS Bedrock, or similar channels will increasingly see reserved-capacity pricing rather than pure pay-as-you-go.

For CFOs, this matters because price declines will now track capital costs and debt servicing schedules, not Moore's Law or competitive dumping. The financing structure mirrors renewable power projects or telecom infrastructure, where pricing stabilizes around cost of capital rather than falling rapidly. Enterprises unwilling to commit to multi-year compute contracts will pay higher premiums for on-demand access, while those signing 12- to 36-month commitments can lock in better rates before the next tranche of financing raises the floor.

Anthropod users should expect more predictable capacity availability as 1 GW of compute comes online, but also tiered service levels that prioritize long-term buyers over ad-hoc usage. The same logic applies to any vendor whose capacity is backed by project finance: reserved access costs less than spot access, and commitments longer than the debt tenor get the best pricing.

Hyperscaler Capex Dwarfs Private Finance

The $35 billion Broadcom-Apollo-Blackstone deal and KKR's $10 billion Helix vehicle represent significant capital, but they are rounding errors against hyperscaler spending. New 2026 capex guidance shows Amazon, Alphabet, Meta, Microsoft, and Oracle together plan to spend $660 billion to $690 billion on AI infrastructure in 2026, nearly doubling 2025 levels.

Amazon projects $200 billion in 2026 capex, most for data centers. Alphabet targets $175 billion to $185 billion, Meta $115 billion to $135 billion, Microsoft at least $120 billion, and Oracle $50 billion. Eighteen months ago, aggregate annual AI infrastructure commitments from these five stood at roughly $380 billion; the increase to $660 billion to $690 billion in 2026 cements hyperscalers as the dominant suppliers of AI compute.

For enterprises, this spending gulf means most production AI workloads will run on hyperscaler infrastructure, but the rise of asset-backed third-party capacity creates meaningful alternatives. A $10 billion fund with NVIDIA as a strategic partner and Vistra handling power reduces the risk that smaller GPU hosting providers run out of money mid-project, making second-tier providers viable for mission-critical workloads.

Energy and Contract Duration Enter the Procurement Conversation

Helix's focus on data centers, power, and network connectivity as a single package signals a new procurement model. Enterprises with Scope 2 or Scope 3 emissions reporting requirements will see more offers where power procurement, sustainability targets, and AI compute capacity are bundled. This is not incidental—Vistra's involvement as the primary power partner means Helix can offer long-term power purchase agreements alongside GPU access, something hyperscalers rarely make transparent.

Contract durations will also shift. Infrastructure funds operate on 10- to 20-year horizons, and they will structure AI campus agreements accordingly. Enterprises accustomed to three-year cloud commits should prepare for vendors backed by project finance to push for longer terms, especially when custom configurations, dedicated capacity, or power guarantees are involved.

What to Watch

Track whether Broadcom's AI XPV platform signs additional anchor tenants beyond Anthropic. If OpenAI, Cohere, or xAI commit capacity under similar financing, the multi-year reserved-capacity model becomes the new standard, and enterprises lose negotiating leverage for short-term contracts.

Watch for hyperscalers to offer their own asset-backed financing products. Amazon, Microsoft, and Google have balance sheets large enough to internalize this model, and they may respond by creating reserved-capacity instruments that look like power purchase agreements—fixed-price, multi-year commitments with penalties for early termination.

Finally, monitor power partner announcements. Helix's exclusive relationship with Vistra gives it access to generation capacity that most competitors lack. Enterprises planning AI deployments in power-constrained regions should evaluate whether their vendors have comparable power partnerships, or risk delays when data center space exists but power connections do not.