AI Usage Is Rising but Spending Is Stalling: How to Build More Cost-Efficient, Secure Applications

What Happened

The latest reporting points to a practical shift in technology adoption: businesses are demanding more useful work from AI without proportionately increasing spending, while autonomous transactions and connected infrastructure create new control requirements.

• AI consumption and spending are diverging. Ramp data indicates that AI usage continues to rise while spending has stalled. Price reductions and tighter controls among major spenders may explain the gap; the reporting does not establish their relative contribution. [2]
• Shopping assistants are gaining transaction authority. Amazon announced texting for Alexa for Shopping. Customers can describe products or send photos, receive recommendations, and have Alexa place orders. The rollout is expected to reach millions of users over the coming weeks, rather than being universally available immediately. [8][10]
• Recommendation design faces regulatory scrutiny. The European Commission issued preliminary findings that TikTok’s recommendation system may breach EU platform rules by creating harmful content “rabbit holes.” These are preliminary findings, not a final judgment. [9]
• Infrastructure security remains configuration-dependent. Reporting from November 5 describes a password-bypass vulnerability involving OpenSSH, musl, and particular account configurations. A separate U.S. government warning describes ongoing attacks affecting connected refrigeration equipment across critical industries. [17][23]

Other developments broaden the picture. Fox announced the close of its $22 billion Roku acquisition, bringing a platform reported to reach more than 100 million U.S. households under its control. Serve Robotics announced connectivity and data services for robotics customers. Neither announcement, by itself, establishes a new general-purpose cloud platform. [5][21]

Why It Matters to Businesses

AI budgets should track outcomes, not request volume. Rising usage with flat spending weakens the assumption that adoption automatically produces higher vendor revenue—or higher customer costs. Buyers should compare cost per successfully completed task, including retries, human review, and downstream corrections, rather than judging value by token prices alone. The Ramp findings support closer scrutiny, not a conclusion that AI demand is declining. [2]

Agentic commerce changes the liability boundary. A recommendation can be ignored; an order creates a financial commitment. Amazon’s announcement illustrates why enterprise agents need explicit authorization, spending limits, and reconciliation when connected to procurement, booking, or payment systems. [10]

Platform dependence includes distribution and policy risk. Fox’s Roku acquisition may affect future distribution negotiations, although the reporting does not establish pricing or access changes. Separately, scrutiny of TikTok shows that ranking behavior—not just prohibited content—can become a compliance concern. [5][9]

Kimbodo Engineering Perspective

Our engineering judgment is to separate three decisions: which model generates a proposal, which policy permits an action, and which service executes it. Combining them in a single prompt makes an application difficult to audit and gives model errors too much authority.

The same separation improves economics. Use smaller models or deterministic code for narrow, measurable tasks, and reserve more capable models for cases that justify their cost. Underdog’s demonstration of a 32MB model generating speech and music is worth watching, but a demonstration without a release date or production benchmarks is not evidence of deployment readiness. [12]

For cloud and developer platforms, optimize for replaceable components rather than unrestricted portability. A thin model-provider interface, centralized identity, and shared evaluation tooling can reduce switching costs. Rebuilding every managed service behind a universal abstraction usually adds complexity before it delivers value.

Finally, recommendation quality needs longitudinal evaluation. A system can produce individually acceptable results while repeatedly steering users toward a narrow or harmful pattern—the type of concern raised in the Commission’s TikTok findings. [9]

How We Would Implement It

Instrument costs and quality first

• Define a successful business outcome for each workflow and build a representative evaluation set before changing models.
• Record model version, tokens, latency, tool calls, retries, review time, and final outcome. Apply tenant-level budgets and concurrency limits.
• Route low-risk tasks to lower-cost execution paths only when evaluations show acceptable quality. Retain escalation paths for ambiguous cases.

Put transaction authority outside the model

• Run application services under separate workload identities. Keep payment credentials and purchasing authority in a restricted execution service, not in model context.
• Require structured action proposals and validate product identifiers, quantities, destinations, and price limits against trusted systems.
• Require user confirmation for purchases initially. Use idempotency keys, durable workflow state, and audit records to prevent duplicate actions and support reconciliation.
• Treat product listings, retrieved pages, images, and messages as untrusted input that cannot override authorization policy.

Harden deployment and device access

• Inventory SSH services, runtime libraries, PAM configuration, and account states. The reported vulnerability requires specific conditions; an Alpine image alone does not establish exposure. [17]
• Where feasible, disable password-based SSH, use short-lived credentials, and restrict administrative access through controlled gateways.
• Isolate operational devices from general business networks, restrict remote access, and monitor configuration changes. For refrigeration systems, preserve local safety controls and manual recovery procedures. [23]

Risks, Costs and Security

Lower inference prices do not guarantee lower total cost. Evaluations, observability, integration maintenance, human review, and incident response remain necessary. Excessive retries or autonomous tool use can erase model-price savings.

More autonomy increases the potential impact of a failure. Transaction limits, approval gates, and restricted identities introduce friction, but constrain mistakes and prompt-injection attacks. Expand autonomy only after measured performance supports it.

Data governance must cover secondary use. Reports that at least 10 local governments have moved to drop Flock surveillance cameras cite privacy, data-sharing, and procurement concerns. Buyers of connected platforms should contractually define retention, onward sharing, access logging, and deletion—not rely solely on dashboard controls. [15]

The near-term priority is not adopting every new model or platform. It is building applications whose cost, authority, data flows, and failure modes remain measurable as vendors and capabilities change.

Sources

1. [2] Ramp data shows AI spending stalling just as Anthropic nears its IPO
2. [5] Fox bought Roku
3. [8] After blocking Muse, Amazon now lets users text Alexa to help them shop
4. [9] TikTok’s For You page needs to change, says the EU
5. [10] Amazon will let you text Alexa your shopping list
6. [12] Underdog shows off a 32MB speech model in new demo
7. [15] At least 10 local governments have moved to drop Flock’s surveillance cameras since summer
8. [17] Meet *sshole, the vulnerability that lets hackers SSH into Linux devices
9. [21] Serve Robotics launches new business to become the AWS for robots
10. [23] US government warns of ‘ongoing cyber threat’ amid mass hacks of industrial retail and military freezers