Payments began this month. Alphabet is sending SpaceX $920 million every 30 days for access to thousands of Nvidia GPUs housed in one of the Colossus data centers. The arrangement, struck in June, runs through mid-2029. Add it up and the total reaches nearly $29 billion. Yet the contract carries exit ramps that could shrink those figures sharply.

SpaceX disclosed the terms ahead of its own public debut earlier this year. The deal followed a larger pact with Anthropic. Together the two agreements promised tens of billions in annual AI infrastructure revenue. But delivery deadlines, termination clauses and Alphabet’s own building plans cast doubt on whether the full amount will ever change hands. The pact reveals more about today’s acute shortages in AI training capacity than any permanent shift in who supplies it.

Google Cloud’s backlog of unsigned revenue stood at $514 billion last quarter. That figure jumped from $460 billion only three months earlier. Demand for its Gemini models, especially the enterprise agent platform, outstripped internal forecasts. A Google Cloud spokesman told The New York Times the agreement supplies “bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected.” Short-term. Timely. Those words matter.

SpaceX built the Colossus cluster originally for xAI and its Grok models. The February merger of xAI into SpaceX turned that internal asset into a commercial one. Anthropic locked up much of Colossus 1 at $1.25 billion per month. Google’s slice covers roughly 110,000 Nvidia GPUs plus supporting CPUs, memory and networking. Capacity ramps through September at lower rates before the full fee kicks in during October.

But the fine print bites. SpaceX must deliver the full GPU count by Sept. 30. Miss that date and, after a one-month grace period, Google can walk away or accept whatever hardware is ready at a reduced price. After Dec. 31 either side can end the contract with just 90 days’ notice. Those provisions appeared in regulatory filings reviewed by Reuters.

Recent reporting suggests the September deadline created real pressure. SpaceX reportedly overhauled data center designs around power and cooling after reliability shortfalls at Tennessee and Mississippi sites. Key infrastructure leaders shifted from rocket programs to fix uptime issues that fell below internal 99.9% targets. As of early October, public updates on full delivery remain absent. The Motley Fool warned investors not to count on the entire sum, noting Alphabet’s massive capital spending will eventually ease its own constraints.

Alphabet already owns roughly 5% of SpaceX. That stake, valued above $80 billion after the IPO, gives the search giant indirect upside if Colossus drives higher valuations. The relationship runs deeper. Google and SpaceX have discussed placing data centers in orbit. Project Suncatcher, Alphabet’s effort to test tensor processing units on satellites, launched its first prototype aboard a SpaceX Transporter mission in early October. The test satellite, built with Planet Labs, carries four TPUs. It will measure radiation tolerance, thermal behavior and laser communications in vacuum.

But. Space-based compute faces steep hurdles. Solar power in the right orbit can reach eight times terrestrial levels. Yet launch costs, radiation hardening and data transmission latency remain formidable. SpaceX itself eyes orbital clusters under concepts like Starmind. Musk has said compute in space will “round up to 100% of all compute” eventually. For now those visions sit years away. The ground-based Colossus deals pay today’s bills.

SpaceX’s AI division posted a $6.4 billion operating loss in 2025 on $3.2 billion revenue. The new contracts flip that math. Annualized revenue from Google and Anthropic alone exceeds $26 billion. Management talks of expanding Colossus further and adding new sites. Still, the customer list shows a pattern. Hyperscalers and AI labs rent capacity while they race to construct their own clusters. Once those plants come online, demand for third-party GPUs could soften.

Google’s capital expenditure runs near $200 billion this year and climbs higher next. Much of that builds data centers and fills them with custom TPUs and Nvidia silicon. The company signaled the SpaceX pact buys time. It does not replace long-term self-sufficiency. Similar dynamics play out across the industry. Microsoft, Amazon and Meta pour hundreds of billions into chips and facilities. The current seller’s market for GPUs won’t last forever.

And yet SpaceX holds advantages. Its vertical integration spans rockets, satellites and now massive power-hungry clusters. Starlink could someday ferry data to orbital nodes. The Memphis supercomputer already operates at gigawatt scale. Recent reports indicate the company scaled from 0.4 gigawatts to 1.4 gigawatts in a single year. Targets reach 2 gigawatts by year-end. Those numbers explain why Alphabet turned to a rival’s infrastructure in the first place.

Executives at both firms stay quiet on exact performance to date. No public disclosure has confirmed whether SpaceX met the September GPU target or triggered any adjustment. The one-month grace window still runs. If Google accepts partial capacity, monthly payments drop proportionally. That would trim the headline $29 billion figure. Even full delivery leaves the 90-day termination option intact after December.

Investors who bought SpaceX shares in the record IPO priced in much of this revenue. The stock reacted positively to launch cadence and new contracts. But sustained AI income depends on renewal or replacement deals once current terms expire. By 2029 both Google and Anthropic expect their internal supply to expand dramatically. New entrants may bid for remaining capacity. Or they may not.

The deeper story is one of temporary arbitrage. Explosive model training demand collided with a multi-year lag in power plant and chip production. SpaceX, through an asset originally meant for Grok, stepped into that gap and extracted rich pricing. Google bought breathing room. The partnership mixes competition and codependence. Alphabet competes with xAI on frontier models while paying rent on its hardware. SpaceX launches Google’s experimental satellites while building rival orbital plans.

Recent X discussions and analyst notes highlight the tension. Some observers question whether orbital data centers can achieve economic scale without thousands of Starship flights. Others see the Google lease as validation that even deep-pocketed hyperscalers need external help today. A CNBC report on the October Suncatcher launch noted Alphabet’s stake in SpaceX now exceeds $82 billion. Shared interest may discourage early termination.

Still, contracts are contracts. The 90-day clause gives both sides flexibility when market conditions change. Power prices, chip availability and model efficiency all evolve quickly. If training one trillion-parameter model becomes dramatically cheaper, the urgency for raw GPU count fades. SpaceX must keep Colossus utilization high and costs low to remain attractive.

Longer term the company bets on vertical control. Rockets lower launch prices. Starlink provides bandwidth. Future orbital clusters could bypass terrestrial power grids and land constraints entirely. Musk’s public comments suggest he views space as the ultimate destination for compute. Google tests the same idea, albeit more cautiously. Their collaboration on the prototype satellite may yield data useful to both.

For industry watchers the Google pact offers a window into AI infrastructure economics. Massive losses at xAI turned into paper revenue once the hardware was opened to outsiders. Pricing per GPU appears higher than some competing deals, reflecting scarcity and perhaps the cachet of Colossus. Yet the exit provisions remind everyone that today’s seller’s market contains tomorrow’s risk of oversupply.

Payments flow now. Questions linger on delivery, uptime and renewal. SpaceX’s stock reflects optimism that AI revenue will compound. Alphabet’s cloud growth depends on securing capacity wherever it appears. The relationship, born of necessity, will face its first real stress test in coming quarters. How both sides handle that test will say much about who truly controls the next phase of artificial intelligence infrastructure.