Technology. Money. People.
28 Sep – 4 Oct 2026
This week’s lead story
Supplier.Financier.Rival.
Google’s Gemini and Anthropic’s Claude compete for customers. Yet Google can benefit from Claude’s growth: it sells computing capacity to Anthropic and owns a stake in the company. Broadcom, meanwhile, has promised financing to a customer that leases its chips.
Google and Anthropic compete, but Anthropic’s growth can also benefit Google. Choose a role to see how.
Supplier. Google sells Anthropic capacity to train and run its models. Claude’s growth could increase those purchases, but not all of its computing runs on Google’s infrastructure.
Shareholder. Google owns a stake in Anthropic. Growth at a rival can also increase the value of that holding.
Rival. Google’s Gemini and Anthropic’s Claude compete in the model market. A shared financial interest does not remove that rivalry.
01 / ANTHROPIC
Buying compute years before it is used.
Anthropic expects to spend at least $518 billion on computing over several years, according to an IPO prospectus seen by Reuters. About 80% of that amount relates to agreements that are non-cancellable or require payment regardless of usage. This is not one year’s expense.
According to a draft IPO prospectus seen by Reuters, Anthropic’s revenue grew roughly twelvefold in 2025 to $4.6bn. Compute costs tripled to $7.33bn, and its operating loss exceeded $8bn. Reuters [1]
The net loss was $42bn, but roughly $34bn came from a non-cash financing item. It does not represent an equivalent cash outflow. Reuters [1]
Anthropic’s compute commitments
Which purchase commitments can be cancelled?
Customers can switch. Payment commitments remain.
Anthropic has committed to buying $221.1 billion of computing capacity from Google and Amazon, roughly €197 billion. It must pay Google an agreed minimum even if it uses less capacity. Its Amazon agreement has a similar provision. The Google commitment runs through 2033 and the Amazon commitment through 2036. Reuters [2]
The payment principle is familiar from a fixed-term gym membership: fewer visits do not reduce the agreed fee. Reserving capacity in advance could help Anthropic grow. Its bills do not automatically shrink, however, if its own customers spend less.
Broadcom ↔ Anthropic
Broadcom finances its own customer.
Broadcom has promised Anthropic up to $42 billion, roughly €37.4 billion, in financing. That would cover about a third of a five-year, $125.2 billion chip lease. The leased TPUs are chips designed for AI computing, developed jointly by Google and Broadcom. The debt could later convert into Anthropic shares.
The financing would help Anthropic pay chip lease costs to Broadcom. Broadcom would be both a hardware supplier and a source of finance for its leasing customer. If Anthropic cannot meet its payments, Broadcom could be affected through both unpaid rent and unpaid loans. According to the prospectus, a default could also accelerate lease payments and restrict access to financing. An obligation to pay does not guarantee that the money will be there. Reuters [3]
Company relationships
Partners can be rivals, too.
Growth can benefit a rival, too.
Google sells Anthropic compute, holds an equity stake and distributes Claude. Its Gemini models compete with Claude. Google’s revenue also depends on whether customers use Claude on its computing services.
Reuters [2]Broadcom
Financing has also been promised to the lease customer.
Broadcom’s financing commitment would help Anthropic pay for leased TPUs. The hardware lease and financing commitment are separate agreements. A commitment is not money already borrowed.
Reuters [3]Amazon
A minimum purchase still applies in quiet periods.
Amazon supplies compute, owns a stake and distributes Claude. Its minimum-purchase provision resembles Google’s: agreed spending cannot freely shrink with usage.
Reuters [2]xAI
Not every contract locks the buyer in.
Reuters reports that Anthropic’s agreements to buy capacity from xAI are largely cancellable on 90 days’ notice. The comparison is between Anthropic’s own arrangements with different suppliers.
Reuters [2]The terms of competition
When a rival’s success serves your interests.
Google has a reason to compete with Claude in the model market. As a compute supplier and shareholder, it may also have a reason to welcome Anthropic’s growth. Broadcom would finance the customer whose lease payments provide its revenue. Rivalry and a shared financial interest can coexist.
An ownership chart alone cannot tell us how competitive a market is. What matters is how many independent routes a customer has to compute, distribution and their own data. Closely connected companies can still compete on price and products.
If partnerships speed up construction and lower costs, customers may benefit. If switching suppliers also requires replacing tools, contracts and working practices, more of that benefit may stay within the supply chain. Evidence of new entrants reaching customers—and customers actually moving their workloads—would tell us more.
Follow the implications
What could happen next?
Explore three questions raised by this week’s news. Compare two possible outcomes for each and the signs to look for. These are possibilities to consider, not forecasts.
How much choice does the customer retain?
Google is a compute supplier, shareholder, distributor and model competitor to Anthropic. Agreements between the same companies can therefore serve several purposes. Source [2]Source [3]
Switching is a workable option
A company can move its work to another provider at reasonable cost and with manageable effort.
It then has a credible alternative to its current contract.
That could strengthen its bargaining position and put pressure on suppliers to improve prices or service.
What would we look for? Actual migrations would offer useful evidence: what did the move cost, how long did it take, and did the service work afterward?
Switching means changing the work
The model, data and workflows are tightly fitted to one provider’s systems.
Moving may require software changes, renewed checks of the results and staff training.
A lower model price alone does not show whether switching is worthwhile.
What would we look for? The comparison also needs the full cost of moving and continuity of the work. The same company may find it easy to replace one tool and difficult to leave another.
Where does the time saved go?
Barclays says Claude sorts and routes messages. Reported usage does not yet measure the outcome of the work. Source [15]
Room for judgement
Less time goes into routine sorting.
People have time to understand unusual cases.
Customers could reach the right resolution in fewer rounds.
What would we look for? Look at resolution time, reopened cases and staff experience. Throughput alone is insufficient.
A harder queue
Easy cases are handled automatically.
Difficult situations accumulate in the human queue.
Work could become more demanding even as the number of cases falls.
What would we look for? This is not a measured Barclays outcome. Assess it through task difficulty, time for judgement and available support.
With more options, who chooses the objective?
GPT-Synopsys aims to connect a model with design tools in an iterative workflow. An engineer evaluates the result. Source [10]Source [11]
Judgement gains scope
The model helps test more designs.
The engineer compares the chips’ energy use, speed and size.
Expertise could become more valuable in choosing between options than in generating them.
What would we look for? Better outcomes across the full design process, including checks and corrections, would support this path.
Defaults become objectives
The tool prioritises an objective, such as chip speed, by default.
Design proceeds faster if the user accepts that priority without examining it.
The toolmaker’s choice could then determine what the design aims to achieve.
What would we look for? Can users change that priority? Can they see how greater speed would affect energy use? Faster computation does not itself help choose the objective.
What stays with us?
Who can still choose differently?
The same ties can accelerate investment and make changing direction harder. This week’s news does not yet settle which effect will prevail. Customer terms, new competitors’ access to the market and who retains the gains from more efficient work would help answer that.
We can want technology to advance and still examine how it distributes power. If a system becomes more efficient, should success also mean that people and businesses find it easier to decide how they use it?
What should we be able to change—the model, the supplier, or the way we work?