AI Next WaveAugust 5, 2026
Big Tech Has Already Signed Up for a Trillion-Dollar AI Bill
Five companies have committed roughly $1.16 trillion to future lease payments, mostly for data centers. The businesses selling power equipment are collecting now, while the returns from all that computing remain a forecast.
The Short Version
The largest technology companies have moved beyond talking about AI demand. They have signed contracts that can last for decades.
Microsoft, Meta, Oracle, Amazon and Alphabet disclosed about $1.09 trillion in lease payments that had yet to begin, mostly for data centers, according to a Reuters analysis of company filings. Meta then signed another $68 billion in data-center leases in July, pushing the known pipeline to roughly $1.16 trillion.
These commitments are disclosed, yet most haven't entered the companies' reported lease liabilities because the facilities aren't ready for use. The payments stretch across many years and can't be treated as an extra trillion dollars of current debt. They still reveal a large fixed wager on future demand for computing.
The bill is creating winners before AI customers prove the final return. Siemens Energy reported record quarterly orders, revenue and profitability as US data centers helped drive gas-turbine demand. Data-center operators and Middle Eastern customers accounted for about half of its quarterly gas-turbine orders.
The World Bank is urging developing countries to take a cheaper route. Its new World Development Report says they can gain from smaller tools adapted to local languages and services without building frontier models or giant computing clusters. It estimates generative AI could meaningfully raise productivity in 16.2% of jobs in developing economies while putting 4.5% at risk of automation. Those are modeled exposures, not observed outcomes.
Two other developments show where the infrastructure reaches ordinary users. A US appeals court allowed Perplexity's agent to keep shopping on Amazon while the underlying case continues. Cybersecurity researchers found a built-in backdoor across more than 20 models of a small-business and home-office router that may be installed in at least 100,000 locations worldwide.
AI is becoming a set of long contracts, power plants, platform permissions and network devices. Each can create value. Each can also lock in cost or risk before the promised result arrives.
The Trillion-Dollar Commitment
The leases start later, but the obligation starts now
Reuters reviewed filings from Microsoft, Meta, Oracle, Amazon and Alphabet and found about $1.09 trillion in future payments under leases that had been signed but had yet to commence. Most relate to data centers needed for AI.
The accounting needs care. A signed lease generally becomes a balance-sheet liability when the facility is available for use. Until then, the company discloses the future payments in notes to its financial statements. The commitments are undiscounted totals spread across many years, while a recognized lease liability reflects the present value of future payments.
That means the $1.09 trillion can't simply be added to corporate debt. The figure is still nearly four times the roughly $285 billion in lease liabilities already recognized by the five companies, according to Reuters.
The exposures also differ. Microsoft disclosed the largest pipeline at $329.1 billion. Oracle reported $260 billion in uncommenced commitments, nearly seven times its recognized lease liabilities. Its data-center leases are generally expected to start from fiscal 2027 through fiscal 2029 and run for 15 to 19 years.
Oracle has warned that the length and pricing of its leases may fail to match its customer contracts. A customer could leave before Oracle's obligation ends. That mismatch turns a forecast about AI demand into a financing risk.
Amazon's $137.21 billion is less comparable because its lease portfolio includes warehouses, offices, aircraft and vehicles. Meta disclosed $278.99 billion, then added $68 billion in data-center leases in July. Alphabet reported $85.2 billion.
These disclosures measure commitments. ROI remains unproven. The companies have secured future capacity. They haven't shown that every site will produce revenue above its power, equipment, financing and lease costs.
Customers should still care. A cloud provider with long fixed obligations has a strong reason to keep utilization high. That can lead to attractive capacity discounts, pressure to bundle AI into existing products or pricing that makes cancellation and data movement expensive.
Buyers can protect themselves by matching the length of a cloud or AI commitment to the evidence behind the workload. A three-year capacity agreement deserves a stable use case, a measured baseline and realistic demand. A pilot that saved employees two hours last month provides weaker evidence than a production service with recurring customer revenue.
Finance teams should separate five claims that often get compressed into one AI return: spending avoided, employee time released, paid demand served, risk reduced and strategic capacity reserved. Only some of those produce cash. A long contract can be sensible for strategic capacity, but the board should know which return it is buying.
Power Sellers Are Collecting
Data-center demand is already changing the energy business
Siemens Energy reported €17.9 billion in orders for its fiscal third quarter, the highest in its history. Revenue rose 18.5% on a comparable basis to €11.4 billion. Profit before special items increased from €497 million a year earlier to €1.62 billion.
The company makes gas turbines, grid equipment, transformers, wind turbines and other systems that sit between a power source and an electricity user. AI data centers have become a material customer.
Chief executive Christian Bruch said data-center operators and Middle Eastern customers accounted for about half of third-quarter gas-turbine orders. Reuters reported that US data-center expansion helped drive the quarter, alongside power projects in the Middle East.
The result shows where hard-dollar AI revenue is easiest to see. Siemens Energy booked equipment and service orders. Its customers are paying for physical capacity. Whether the applications running inside those data centers earn an adequate return remains a separate question.
There is also an energy consequence. The demand is lifting gas-turbine orders at the same time companies and governments pursue emissions goals. Gas plants can be built to balance grids and replace higher-emitting generation, yet they create long-lived fuel and carbon exposure. A server may be replaced in a few years. A turbine can shape a regional power system for decades.
Communities evaluating a data-center project need a complete bill. The useful questions cover new generation, transmission, backup systems, water, local tax incentives and who pays when demand arrives faster than the grid can serve it. A promise that the facility will eventually use clean power should include dates, contracts and the emissions created before that supply exists.
The workforce opportunity is substantial and physical. Turbine service, transformer manufacturing, substation work, power electronics, controls, cooling and high-voltage maintenance all need trained people. A local college or apprenticeship program should build around confirmed project schedules, required credentials and equipment that employers actually use.
For smaller businesses, the practical opening lies in the supporting work. Electrical inspection, thermal monitoring, backup-power testing, water treatment and maintenance documentation can become recurring services. A credible offer starts with one asset class and earns its case through avoided downtime, faster response and verified compliance.
Siemens Energy's quarter is reported performance. The durability of the data-center cycle is still a forecast. Delayed sites, constrained grids, weaker AI demand or customers shifting toward more efficient systems could change the order path.
A Cheaper Route to Useful AI
The World Bank says developing countries should begin with the problem
The World Bank's World Development Report 2026 makes a deliberately different argument from the trillion-dollar buildout.
Developing economies can use existing, lower-cost AI tools and adapt them to local languages, institutions and delivery channels. They don't need to reproduce a frontier model or a hyperscale data center to improve a health workflow, help a farmer interpret weather information or give a teacher better materials.
The report estimates that generative AI could meaningfully boost productivity in 16.2% of jobs in developing economies, close to 18.7% in high-income countries. It puts 4.5% of jobs in low- and middle-income countries at risk of automation, compared with 14.2% in high-income countries.
Those percentages describe modeled exposure based on the tasks inside jobs. They aren't a forecast that exactly 4.5% of workers will lose employment or that every exposed worker will become 16.2% more productive. Actual results will depend on adoption, wages, business formation, worker bargaining power and whether institutions can use the tools safely.
The basics remain uneven. Nearly one-third of rural schools in Sub-Saharan Africa lack reliable electricity, and more than two-thirds lack dependable internet access, according to the World Bank. A language model can't solve an access problem when a school can't keep a device connected.
The report's strongest advice is sequencing. Adopt tools that already exist. Adapt them to local data, language and institutions. Advance toward more complex development as skills and infrastructure grow.
That sequence also works for a small business or public agency in a wealthy country. Start with the service failure. Choose a tool that fits the available data and the people who will operate it. Test whether the result improves access, speed, quality or cost. Build custom technology only when the existing options fail an important requirement.
A health clinic could pilot AI-assisted intake for one narrow category of appointment. The prerequisites are a reliable record system, consent rules, a route for urgent cases and staff who can correct mistakes. Useful measures include wait time, missed urgent cases, correction rates and whether patients with limited literacy or language access receive equal service.
The World Bank also warns that imported systems can concentrate market power, weaken privacy and increase dependence on a small number of providers. Local adaptation needs technical skill and institutional judgment. Translating the interface without evaluating the recommendations can make a weak system easier to use.
Public procurement should demand observed outcomes. Count people served, decisions corrected, time to resolution and cost per successful case. Vendor forecasts and demonstration videos can justify a test. They can't prove the return.
When the Customer Sends a Bot
A court put the user between the agent and the platform
The Ninth US Circuit Court of Appeals vacated a preliminary injunction that had blocked Perplexity's shopping agent from operating on Amazon. The underlying lawsuit continues.
Amazon argued that Perplexity violated federal and California computer-access laws after its Comet browser entered password-protected customer accounts. Perplexity said the user authorized the agent and that Amazon was trying to protect its control over shopping.
The appeals court focused on who accessed Amazon's computers. It concluded that the user accessed Amazon with help from Perplexity's agent. Perplexity's servers sent instructions after receiving screenshots from the browser, but the company's computers didn't directly enter Amazon's systems.
That distinction was enough to make Amazon unlikely to succeed on its computer-access claims at this preliminary stage. The court vacated the injunction and returned the case for further proceedings. It didn't give every shopping bot a general legal right to ignore a retailer's rules, settle the security dispute or end Amazon's lawsuit.
The commercial consequence is immediate. Agents can compare products, navigate accounts and place orders while bypassing parts of the interface a retailer designed. That can reduce friction for a customer and reduce the retailer's control over advertising, recommendations, data collection and product presentation.
For shoppers, delegation creates a new verification problem. The agent may use account credentials, send screenshots to an outside server and make a purchase based on incomplete product information. A useful setup limits the budget and product category, requires approval before payment and preserves a clear record of the sources and actions.
Retailers need an agent policy that separates human-authorized automation from credential theft, scraping and abusive traffic. Blocking every automated tool can hurt accessibility and customer choice. Accepting opaque agents can expose account data and create disputes over returns, fraud and consent.
A practical pilot can begin with read-only product search. Require the agent to identify itself, show which data leaves the device and ask for confirmation before it changes an account or spends money. Measure task completion, incorrect purchases, account lockouts, customer complaints and support time.
The court ruling moves one legal boundary. Trust will still depend on the product design around permission, disclosure and correction.
The Backdoor in the Small-Office Router
More than 20 models may be waiting for commands
Cybersecurity firm VulnCheck says more than 20 router models sold under the Zbtlink and Wiflyer names ship with a built-in remote-control implant. The devices are marketed for homes, small offices, vehicles and other locations that need cellular connectivity.
The implant, which VulnCheck calls ENDLESSDOORS, runs with full administrative privileges. It contacts a specific internet address and a Chinese-registered domain about every 35 seconds. Whoever controls the destination can send a command or request an interactive root shell, according to the researchers.
The router starts the connection from inside the network. A normal firewall or network-address translation rule that blocks unsolicited inbound traffic may therefore provide little protection.
VulnCheck demonstrated control of a purchased AX3000 unit in an isolated lab. Its chief technology officer estimates that at least 100,000 affected routers are deployed worldwide. That deployment figure is an estimate. Reuters couldn't determine why the code was installed, whether the control infrastructure has been used against customers or how many active devices sit in the United States. Zbtlink didn't respond to Reuters.
The design remains unacceptable even without proven abuse. A hidden root-level control channel creates a route to every system that trusts the router. A clinic, construction crew, retail site or mobile office may use one because it is inexpensive and easy to deploy.
Owners should identify the exact maker and model on the physical label and purchasing record. If it matches an affected model, isolate it from sensitive devices and replace it with equipment whose firmware and support path can be verified. A factory reset may reload the same code because the implant ships with the device.
Managed-service providers can look for repeated outbound connections to the indicators published by VulnCheck, yet detection shouldn't become a promise that a compromised router is safe. The durable fix is removal, credential rotation and review of other devices for unusual access.
This incident also belongs in AI procurement. New cameras, sensors, robots and edge-computing systems can arrive with embedded routers and white-labeled network equipment. A risk review that stops at the logo on the finished product can miss the component that controls every connection.
Opportunity Radar
AI commitment ledgers for growing organizations
Cloud resellers, software companies, hospitals and larger professional firms can accumulate reserved compute, minimum-spend agreements, bundled assistants and long software terms across separate budgets. Few have one view of the obligation, actual use and revenue or service outcome supported by it.
An operations and finance consultancy could build a commitment ledger for one organization. The service would match contracts with usage, cancellation terms, workflow owners and measured benefits. The buyer gains a renewal decision and a view of stranded capacity. The provider must validate invoices and contract language, then separate cash savings from released employee time. A dashboard without a decision deadline or accountable owner becomes another subscription.
Embedded network-device reviews
Small manufacturers, clinics, schools, construction companies and local governments often buy connected equipment through an installer and never receive a complete inventory of the networking components inside it.
A cybersecurity firm, electrical contractor or managed-service provider could review one device class. The work would verify physical models, firmware sources, outbound connections, vendor support and replacement procedures. Customers benefit through lower intrusion and downtime risk. Validation requires inspection on site and controlled traffic monitoring. A vendor questionnaire alone can miss a white-labeled component.
What You Can Do With This
If you approve AI spending
Place every long-term commitment beside the workload, baseline and expected return. Record which benefit is cash, capacity, quality, risk reduction or strategic learning. Set a review date before the cancellation window closes.
If your community is courting a data center
Ask who funds generation, transmission, water and backup systems. Request the energy mix and emissions by year, along with hiring requirements, tax incentives and protections for existing ratepayers.
If you deploy an agent
Begin with read-only access and a narrow task. Limit credentials, budget and destinations. Require approval before purchases or account changes, and keep a record that a person can review after an error.
If you operate a small network
Inventory the actual router model, including devices embedded in other equipment. Review outbound traffic, replace unsupported or affected hardware and rotate credentials after removal.
The Bigger Picture
The AI buildout is producing fixed commitments before it produces fixed answers about return.
Five technology companies have signed years of future lease payments. Siemens Energy is booking orders for the machinery that supplies the power. Developing countries are being told to capture value with smaller tools and stronger foundations. Shopping agents are testing who controls a customer's path through a platform. A low-cost router shows how one hidden component can compromise the whole environment.
Scale changes the cost of being wrong. An abandoned prompt experiment wastes minutes. An unused enterprise license wastes a renewal. An underused data center can leave a 15-year obligation. A new gas plant shapes emissions and customer bills. A vulnerable router can expose every trusted device behind it.
The response is practical discipline. Match the term of the commitment to the quality of the evidence. Build a small pilot before a large obligation. Measure the customer or public outcome after the demonstration ends. Inspect the infrastructure that carries the system, including the parts sold under another company's name.
AI can expand capacity, improve access and create new services. Its value will be judged against the contracts, power systems, permissions and security decisions that remain after the excitement moves on.
References
Reuters: Big Tech's uncommenced leases and the roughly $1.16 trillion known pipeline, August 4, 2026
Siemens Energy: Fiscal third-quarter orders, revenue, profit and outlook, August 5, 2026
Reuters: Data-center demand and Siemens Energy's record quarter, August 5, 2026
World Bank: World Development Report 2026 overview and main messages
World Bank: Jobs, infrastructure gaps and policy recommendations from the new report, August 4, 2026
Reuters: World Bank findings on AI, jobs and developing economies, August 4, 2026
Ninth US Circuit Court of Appeals: Amazon.com Services v. Perplexity AI opinion, August 4, 2026
VulnCheck: Technical analysis and affected-model indicators for ENDLESSDOORS, August 5, 2026
Reuters: Researchers document a backdoor in Zbtlink and Wiflyer routers, August 5, 2026