July 21, 2026

AI's Hidden Bill Is Showing Up in Your Inbox, Your Home and the Power Grid

Deepfake scams are scaling across borders. Tesla's autonomy bet is consuming billions. China may restrict access to its models, while household batteries and thermostats are being recruited to support the grid.

By Jay Cadmus

AI Next Wave Is Back

I started AI Next Wave to make sense of artificial intelligence before it became a label pasted onto nearly every product, pitch and corporate plan. The publication is returning with a wider lens.

AI will remain at the center because it's changing how people work, build, create, learn and compete. We'll also follow the technologies and systems shaping what AI can actually do: energy, cybersecurity, robotics, chips, digital identity, mobility, healthcare, policy and the infrastructure behind the screen.

The goal is useful context. Each edition will look for developments that could change a profession, create an opening for a business, affect a household, reshape an industry or introduce a risk worth preparing for.

Today's stories are a good place to restart.

The Short Version

AI has become easier to use, easier to fund and easier to misuse. The costs are appearing in places far beyond a software subscription.

Criminal networks are using synthetic identities, translated messages and deepfake calls to run scams across languages and borders. Tesla is expected to burn cash as its spending on AI, robotics and manufacturing climbs toward $25 billion this year. Washington and Beijing are preparing for formal AI talks while China considers controls that could limit overseas access to advanced models and chip technology.

The physical cost is moving closer to home. Utilities and technology companies are connecting batteries, electric vehicles and thermostats into virtual power plants that can help accommodate data-center demand. France is preparing to vote on a social-media restriction for children under 15, pulling age verification and digital identity deeper into everyday online life.

These developments carry practical consequences. People need stronger ways to verify who's contacting them. Companies need evidence before funding an AI project at scale. Builders need alternatives in case politics changes access to a model. Households need to understand what they're giving up when a utility controls a connected device. Age checks need to protect children without building another warehouse of sensitive identity data.

Digital Trust

AI Has Become Part of the Scammer's Operating System

A United Nations Office on Drugs and Crime report released Tuesday estimates that scams connected to Southeast Asia's expanding criminal ecosystem caused between $88.3 billion and $114.1 billion in combined losses during 2025.

That wide range deserves attention. Scam losses are difficult to measure because many victims never report them and countries track fraud differently. The UN figure is an estimate, not an audited total. Its size still shows the scale of the problem.

The organizations behind these schemes combine cryptocurrency, encrypted messaging, malware and satellite internet with generative AI. They can create convincing phishing messages, manufacture identities and conduct deepfake video or voice calls in real time. AI-powered translation lets one operation target several languages at once.

Some scam compounds rely on people who were trafficked and forced to work. Enforcement raids have disrupted several large sites, but UN researchers say operators often move into smaller properties, neighboring countries or harder-to-trace online infrastructure.

The familiar warning signs are losing value. Clean grammar and a recognizable voice no longer prove that a message came from someone you know. A live-looking video call can be manipulated too.

Families can create a private verification phrase and agree that every urgent request for money gets confirmed through a known phone number. Businesses can require a second person to approve payments and use a separate channel to verify changes to payroll, bank details or vendor accounts. The person asking for the payment shouldn't control the verification channel.

The same rule applies to AI agents. An agent may be allowed to read an invoice, compare it with a purchase order and prepare a payment for review. Authority to change account information or release money creates a much larger target. A person should remain responsible for that final step.

Cybersecurity consultants, accountants and managed-service providers have a credible service opportunity here. Small businesses, schools, nonprofits and local governments need fraud drills built around cloned voices, fake executives and vendor impersonation. The exercise should measure whether employees use the verification process under pressure, how quickly they report the attempt and whether the controls block a fraudulent payment without delaying normal work.

Work, Money & Opportunity

Tesla's AI Plan Is Reaching the Expensive Part

Tesla is scheduled to report second-quarter results Wednesday. Analysts surveyed by LSEG expect the company to post about $3.3 billion in negative free cash flow, which would mark its first quarterly cash burn in over two years.

That figure is a forecast. Tesla hasn't reported the quarter, and the result could differ.

The spending behind it is real. Tesla has raised its projected 2026 capital spending to roughly $25 billion as it funds AI infrastructure, data centers, manufacturing, robotaxis, the Cybercab and its Optimus humanoid robot.

Tesla's robotaxi service currently operates in Austin, Dallas, Houston and Miami. The company has started manufacturing the Cybercab, but the purpose-built vehicle hasn't entered the network. Elon Musk has said its production ramp will be "agonizingly slow."

Tesla's vehicle business delivered a record April-to-June quarter and may generate more cash if the recovery continues. Analysts still expect automotive margins to tighten compared with the previous quarter.

Wednesday's report can reveal the current financial strain. The commercial outcome of Tesla's autonomy and robotics investments will take much longer to establish.

Companies funding their own AI programs face a smaller version of the same discipline problem. Spending shows commitment while leaving customer demand, technical readiness and profit unanswered. Each funding stage should require stronger evidence.

An autonomy program might begin with safe miles and the frequency of human intervention, then add vehicle availability, paid rides, insurance costs and revenue per vehicle. A workplace agent might begin with correction rates and time saved, then add sustained usage, completed transactions, support cost and the quality of the finished work.

Time savings create capacity. Financial return appears when the organization uses that capacity to reduce cost, increase output, improve quality, lower risk or earn revenue. Leaders should name which form of value they expect and collect a baseline before the pilot begins.

Strategic learning has value as well. It still needs a spending limit, an owner and a decision date.

Policy, Power & Access

Model Access Is Becoming Part of Trade Policy

The United States and China are planning formal AI discussions for September, Reuters reported Tuesday, citing five people familiar with the matter. Dates and participants haven't been finalized, and neither government has publicly announced the meeting.

The talks may cover definitions of frontier AI and risks involving military systems, cyberattacks and labor. A first session would probably focus on basic terms and working relationships because the two countries remain far apart on chips, models and national security.

China is separately considering tighter export controls on AI models, training data and semiconductor technology, according to Financial Times reporting. Regulators have reportedly consulted Alibaba, ByteDance, Zhipu and other technology companies. The proposals remain under review, and Reuters said it couldn't independently verify the details.

The direction is still useful to watch. Governments increasingly treat model weights, training knowledge and chip designs as strategic assets. China has promoted open models as an affordable option for international users, especially in developing economies. Future controls could limit access to the most advanced technology or restrict how it moves across borders.

Builders and businesses should plan for access to change. A model available in one market today could face licensing, export or hosting restrictions later.

Portability begins with ordinary technical choices. Keep prompts, evaluations and business rules outside a single vendor's interface when possible. Test an important workflow with a second approved model. Record where the data is stored and whether each provider can operate in the countries you serve.

A backup model sitting unused on a procurement list provides little protection. Teams need to know how it performs on their own work and how long a switch would take.

The policy shift also creates demand for specialists who can translate trade rules, data-residency requirements and vendor dependencies into practical product decisions. Large multinationals already buy this expertise. Smaller software firms, exporters and consultancies will need a version sized for their own contracts and markets.

Infrastructure Behind the Screen

Your Thermostat May Help Make Room for a Data Center

Data centers need large amounts of electricity, while new power plants and transmission lines can take years to build. Utilities and technology companies are looking for flexibility inside equipment that already exists.

A virtual power plant connects many small energy resources such as home batteries, electric vehicles, solar systems, water heaters and smart thermostats. Software can reduce demand or send stored electricity to the grid during a peak. Thousands of devices can then operate as one flexible resource.

Wood Mackenzie's 2025 market report put North American virtual-power-plant capacity at 37.5 gigawatts, up 13.7% year over year. The US Department of Energy has estimated that 80 to 160 gigawatts by 2030 could cover 10% to 20% of peak demand and avoid about $10 billion in annual grid costs. Those figures describe modeled potential, not guaranteed savings.

Google and Voltus announced a three-year agreement on June 2 to aggregate as much as 100 megawatts of flexible capacity across the PJM power market. Google will fund the program, while participating homes and businesses can receive payments through Voltus.

The attraction is speed. Adjusting a thermostat or discharging a battery can provide capacity much sooner than constructing a new plant. Customers may earn money for allowing limited control of participating devices.

The household economics will vary. Equipment cost, financing, local program rules, battery wear and the frequency of grid events all affect the result. Vendor claims about savings should be checked against the customer's final bill.

A 2023 Brattle Group study prepared for Google modeled 400 megawatts of resource adequacy from virtual power plants and compared it with conventional alternatives. The study estimated that broader VPP deployment could save US utilities $15 billion to $35 billion in capacity investment over ten years. These are modeled savings from a sponsored study, not audited results from a nationwide program.

Consumers should ask how often a utility or aggregator can adjust a device, what battery reserve remains during an outage, who can see usage data and whether leaving the program carries a fee. Contractors can help customers evaluate those terms.

Electricians, HVAC companies, solar installers and energy consultants already sit close to the decision. They can add virtual-power-plant readiness to an energy audit, confirm equipment compatibility and compare the incentive with the customer's real operating cost.

Virtual power plants can reduce some peak demand and defer certain upgrades. Grid planners will still need new generation, storage and transmission as data-center demand grows.

Life Online

France Is Preparing to Test Whether a Social-Media Age Limit Can Work

French lawmakers are scheduled to vote Tuesday on legislation intended to prevent children under 15 from opening social-media accounts beginning September 1. Platforms would receive four additional months to close existing underage accounts. Educational sites and online encyclopedias would remain available.

As of Tuesday morning, the measure hadn't passed its final parliamentary votes. The European Commission will also assess whether the French approach complies with EU law. Those steps could affect the final language and implementation date.

The proposal includes age-verification systems approved by France's privacy regulator and would extend phone restrictions into high schools.

The policy follows serious concern about how social platforms affect young people. France's health and safety agency, ANSES, reports that half of French adolescents spend between two and five hours per day on a smartphone and 58% of people ages 12 to 17 use social media daily. Its review of over 1,000 studies found risks involving sleep, self-esteem, anxiety, depression, harassment and exposure to harmful content. ANSES recommends safer platform design alongside reliable age checks and digital education.

Age verification introduces a separate privacy problem. A system may request an identity document, estimate age from a face or rely on confirmation from another account. Poor design can create a database connecting a child's identity with online activity. Better systems return a limited age signal and retain as little personal information as possible.

Enforcement will be uneven. Some children will try borrowed accounts, virtual private networks or smaller services. Parents and schools will still need to discuss manipulation, harassment, privacy and what to do when harmful content appears.

The legislation can change access rules. Safer product design, parental involvement and digital literacy still determine much of what happens after a child picks up a phone.

Age assurance is becoming a serious market for identity companies, privacy engineers and child-safety specialists. Providers will need to prove that their systems work across different faces and devices, resist spoofing, minimize data collection and give users a way to appeal mistakes.

Opportunity Radar

Deepfake Fraud Readiness for Small Organizations

Accountants, cybersecurity firms and operations consultants can build a practical service around payment impersonation. A convincing voice or video can pressure an employee into changing bank details or sending money.

A credible engagement would map payment authority, establish separate-channel verification and run a simulated incident. The buyer could be a small business, nonprofit, school, local government or insurer. The test should reveal whether employees follow the process under pressure and whether the new controls slow legitimate payments.

Energy-Flexibility Guidance for Homes and Local Businesses

Contractors installing batteries, EV chargers, solar panels and smart thermostats can help customers evaluate virtual-power-plant programs. Customers may benefit from lower bills, participation payments and better backup-power planning.

The service needs an active local program, compatible equipment and transparent calculations. Actual savings after hardware, financing and maintenance will determine whether customers see value.

What You Can Do With This

If You Handle Money

Require a callback to a known number for changes to bank, payroll or vendor information. Add a second approval for payments and test the process with a realistic impersonation attempt.

If You're Funding an AI Project

Name the next piece of evidence that would justify more spending. Use a business outcome such as paid usage, lower correction cost, faster completion or fewer errors. Set the review date and spending limit before the project begins.

If You Build With AI

Run one important workflow on a second model this month. Record quality, cost, response time, data location and migration effort. You'll find out whether the backup can carry real work.

If You're a Parent or Educator

Create a family verification phrase for urgent calls and explain why a familiar voice can be faked. Review social accounts together and agree on how children should respond to strangers, requests for images and content that makes them feel unsafe.

The Bigger Picture

Technology becomes consequential when it connects with the systems people depend on.

AI-generated media now connects directly to payment fraud. Tesla's AI ambitions connect to factories, vehicles and a large capital budget. Models and chips connect to trade policy. Data centers connect to household energy devices. Social-media access connects to age verification and identity.

Every connection creates value and another point where something can fail. A deepfake becomes dangerous when it can trigger a payment. An autonomy program becomes financially dangerous when spending outruns evidence. A low-cost model becomes a dependency when access changes. An age check becomes a privacy risk when it collects too much.

AI Next Wave will follow these connections. The most useful technology coverage helps people decide what to use, what to build, what to fund, what to question and where an opportunity may be forming.

That's where we're starting again.

References

• Associated Press: UN report details the global expansion of Southeast Asian criminal networks and AI-enabled scams, July 21, 2026
• United Nations Office on Drugs and Crime: How AI and digital services are changing fraud operations
• Reuters: Tesla's expected cash burn tests investor faith in its AI and robotics spending, July 21, 2026
• Tesla Investor Relations: Second-quarter 2026 production, deliveries and earnings schedule
• Reuters: Tesla increases its 2026 capital-spending plans for AI, robotics and manufacturing, April 22, 2026
• Reuters: The United States and China plan formal AI talks for September, July 21, 2026
• Financial Times: China weighs tighter export controls on AI models and chips, July 21, 2026
• Wood Mackenzie: North American virtual-power-plant capacity reaches 37.5 gigawatts
• US Department of Energy: Virtual power plants, peak demand and potential grid savings
• Voltus: Google agreement will aggregate up to 100 megawatts of flexible capacity
• The Brattle Group: Modeled capacity-investment savings from virtual power plants
• Reuters: French lawmakers prepare to vote on an under-15 social-media restriction, July 21, 2026
• ANSES: Scientific review of social-media use and adolescent health
• French Senate: Legislative record for protecting minors from social-media risks