AI Spending and ROI: What This Week’s Earnings Reveal
Big Tech earnings show where AI money is going, but not whether customers benefit. Learn how to read capex, credit risk and real-world AI returns.
The AI Spending Test: Who Is Actually Getting a Return?
Big Tech earnings this week should give investors a clearer look at the AI buildout. Microsoft and Alphabet report after the Federal Reserve's Wednesday rate decision, with Amazon due Thursday. Investors will focus on cloud growth and guidance on data centers, chips, and other infrastructure. But strong sales from AI providers don't automatically mean the companies using AI are actually making money.
That gap is where things get interesting - and risky. Cloud and model suppliers can record revenue as customers experiment, even if those customers haven't found a repeatable payoff yet. At the same time, lenders are already pricing in stress in parts of the AI infrastructure market. Reported insurance costs on CoreWeave debt have reportedly moved above levels often associated with distress.
This article lays out a practical way to read the week's signals. We'll separate supplier revenue from customer value, explain what capex and credit prices can (and can't) tell you, and then look at the operational and social risks that sit next to the financial story. The goal isn't to guess whether AI will pay off eventually. It's to spot what evidence would show it is paying off now.
Why AI spending and AI returns are different questions
The key distinction is simple: a company selling AI capacity doesn't prove that a customer is getting a return from using it. Cloud providers can report more AI-related demand while customers are still running pilots, paying for capacity they don't fully use yet, or struggling to make model outputs fit real workflows.
That's why you have to read results from both sides of the deal. Microsoft, Alphabet, and Amazon can show how much demand they see and how much they're investing to meet it. But you still need customer outcomes and operating decisions to judge whether the spending is productive.
Revenue is a signal, not a verdict
AI revenue matters because it means someone is willing to pay. It doesn't tell you who's paying, how consistently they use the product, whether the provider can earn profits, or whether the customer saves more than the tool costs.
The same caution applies to adoption numbers. A company can have thousands of employees with access to an assistant. Access isn't the same as steady use, and steady use isn't automatically business value. The questions to ask are more concrete: Did the tool cut time on a specific task? Did output quality hold up? Did the business avoid adding labor - or increase output - without adding comparable cost?
The value is unevenly distributed
Apollo chief economist Torsten Slok has pointed to a widening gap in profit margins: the seven largest technology companies reportedly saw margins rise from around 15% to around 25% between early 2023 and early 2026, while the other 493 companies in the S&P 500 stayed near 10%. That comparison doesn't prove AI caused the shift. Still, it gets at the central problem: gains at infrastructure suppliers don't automatically spread through the rest of the economy.
Returns may also show up at different speeds. A cloud provider can sell capacity as soon as it's deployed. A customer might need months to integrate a model, change processes, train staff, and manage errors. If that integration doesn't improve a measurable outcome, the customer may reduce or cancel spending - even while the supplier looks strong in the current quarter.
The practical test: treat supplier growth as evidence of demand, then look separately for signs of customer productivity, revenue gains, or cost reductions. Supplier growth can't stand in for customer ROI.
What to look for in Big Tech earnings
For Microsoft, Alphabet, and Amazon, headline AI growth is only part of the story. The real details are how investment, capacity, and monetisation line up over time. A company can look like it's growing quickly while still taking on risk - if spending rises faster than useful demand, or if new infrastructure takes too long to convert into revenue.
Read capital expenditure alongside cloud growth
Capital expenditure, or capex, shows how much a company is putting into long-lived assets like data centers and servers. It isn't a direct measure of AI success. Higher capex can reflect confidence in demand, but it can also reflect the cost of staying competitive - or building capacity earlier than demand arrives.
So read capex alongside cloud revenue growth, management's comments on capacity, and when returns are expected. If investment jumps, ask whether cloud growth is accelerating as well, whether supply constraints are easing, and whether the company expects new capacity to generate returns quickly - or only over a longer stretch.
A useful comparison isn't "more capex is bad" or "more AI revenue is good." It's how much extra investment is needed for each extra dollar of revenue, and how dependable is that revenue. Public earnings usually don't break down AI returns cleanly by project, so you often have to look for consistent signals across multiple quarters.
Separate committed demand from trial usage
Some AI demand is experimentation: teams test features, developers explore model APIs, and customers run short pilots. That can be real demand - but it's not the same as durable usage. Stronger signs include recurring use, renewals, expanding workloads, and customers embedding AI into processes they rely on.
Listen for specifics, not just general enthusiasm. Management may cite customer interest or growing workloads. Those points help - but they're not the same as showing retention, margins, or the economic value customers are getting. When those details aren't available, keep the uncertainty visible instead of treating a confident forecast as proof.
Include the financing environment
The Federal Reserve's rate decision matters because this infrastructure isn't cheap, and a lot of it is financed over time. Higher borrowing costs can make long payback periods less attractive - especially for companies without the balance sheets of the biggest cloud providers. A rate decision isn't a judgment on AI demand, but it does change the cost of building.
For business leaders making decisions inside their own companies, the same discipline applies. Compare the full cost of deployment - not only the model subscription - with the expected benefit. Include integration, security review, employee time, monitoring, and ongoing usage. If a project only looks attractive when you ignore those costs, the business case isn't ready.
Why AI infrastructure debt deserves attention
The firms building and renting AI infrastructure don't all have the same financial resilience. Some operate at massive scale with established cloud businesses. Others borrow heavily to acquire or lease compute, then rely on customers and financing markets to keep the equipment running. That difference matters when demand shifts or funding gets more expensive.
CNBC reported that insurance against a CoreWeave default reached about 1,082 basis points. Credit default swaps are contracts that offer protection if a borrower defaults; a higher price generally reflects higher perceived risk. Levels above 1,000 basis points are often treated as a distress signal. It's a market price, not proof that a company will default.
Why credit signals can move before the business does
Credit markets can reprice risk quickly. Investors may get nervous about debt levels, customer concentration, refinancing needs, or even the uncertain resale value of specialized equipment - long before those concerns show up as missed payments. That makes credit prices worth watching, but they should be read alongside cash flow, debt maturities, customer commitments, and how much equipment is actually being used.
It's also worth looking at broader interconnections. CNBC reported that Nvidia had about $7 billion invested directly in AI cloud providers and $164.5 billion in guarantees and commitments, for a combined $171.5 billion. That doesn't mean the entire amount is immediately at risk. It does suggest how tightly chip suppliers and infrastructure providers are linked - and why a slowdown can spread across multiple parts of the ecosystem.
Ask who carries the downside
When a provider builds capacity using borrowed money, it carries the risk that demand, pricing, or equipment utilisation won't meet expectations. Customers face a different risk: locking into long-term contracts before they know whether their workloads will deliver real value. Chip suppliers can benefit from equipment sales even while other participants absorb more of the financing and utilisation risk.
That's why share prices alone are an incomplete guide. Shares of CoreWeave, Nebius, and IREN were each reported down at least 30% during the month described in the source coverage. These moves can reflect changing expectations, not only current results. For a more grounded view, track three things together:
- Funding: How much debt is due, and can it be refinanced on reasonable terms?
- Utilisation: Is installed capacity being used and paid for?
- Customer quality: Are contracts recurring and diversified, or tied to a few large buyers?
The point isn't to avoid AI infrastructure investment. It's to distinguish a strong long-term demand story from a financing structure that only works if growth stays unusually fast.
How businesses can test AI returns
Businesses don't have to wait for the market to settle before making smart AI decisions. They can make the ROI question concrete by starting with a specific workflow, a baseline, and a stopping rule. That's more reliable than buying broad access and waiting for "productivity" to appear later.
Start with a costly, repeatable task
Pick a task that comes up often and has an observable cost or outcome: support-ticket triage, document review, software testing, or internal knowledge search. Skip vague goals like "make the company more innovative." That might sound good, but it's hard to measure and easy to use as cover for spending without accountability.
Document the current process before you introduce a model. Track time per task, throughput, error rates, rework, and any downstream impact that matters. Then run a controlled pilot with a clear comparison. For instance, compare AI-assisted support responses with the existing workflow, including escalation rates and customer satisfaction - not just drafting speed.
Calculate total cost and the cost of failure
A solid business case includes more than the subscription or API bill. Add integration work, data preparation, security controls, staff training, human review, monitoring, and the cost of fixing errors. If model usage varies, estimate costs under realistic and high-volume scenarios, not just the small pilot run.
Then measure the outcome that matters. If your goal is lower handling time, check whether the time saved actually reduces costs or lets the team handle more. If your goal is faster software delivery, verify that quality and maintenance stay stable. A tool that produces drafts faster but increases review work may just shift labor instead of removing it.
Use a decision rule before the pilot begins
Set a review date and decide in advance what would justify expanding, changing, or stopping the test. A basic decision framework can look like this:
- Expand if the tool improves the chosen outcome, stays within the full cost limit, and meets quality and safety requirements.
- Adjust if the workflow shows promise but needs a different model, narrower scope, or stronger human review.
- Stop if measurable benefits don't cover the cost, if quality drops, or if the process creates unacceptable privacy or compliance exposure.
This helps prevent sunk-cost thinking from taking over. It also gives teams evidence even when a pilot fails - whether that means the model isn't good enough, the workflow was poorly chosen, or the underlying bottleneck still hasn't been addressed.
More elaborate coding agents can be a useful reminder to test efficiency rather than assume complexity pays. A study by Shandong University, as reported by VentureBeat, compared Claude Code, OpenCode, and mini-SWE-agent across five models and nearly 500 SWE-Bench Verified tasks. The reported success rates were statistically similar, while the most elaborate setup cost up to three times as much per task. The real question isn't which tool looks most sophisticated. It's which setup delivers an acceptable result at a cost you can sustain in your own environment.
Key takeaways
- AI supplier revenue doesn't prove customer ROI. Look for evidence from the organizations paying for and using the tools.
- Read capex alongside cloud growth and guidance. Spending alone can't show whether infrastructure is earning a return.
- Treat credit prices as a warning signal, not a prediction. Check debt, refinancing needs, utilisation, and customer concentration together.
- Measure AI in a defined workflow. Include integration, review, monitoring, and error costs in the business case.
- Trust and deployment conditions affect returns. Provenance, safety, energy access, and community acceptance can determine whether an AI system works at scale.
Conclusion: Follow the value, not the spending
This week's earnings may suggest the biggest tech companies still see strong demand for AI and are willing to invest heavily to serve it. That's important, but it answers only the supply-side question. It doesn't settle whether customers can turn AI into durable productivity - or whether every infrastructure borrower can finance its buildout safely.
A better way to read the AI boom is to follow the chain: capital spending, infrastructure utilisation, customer adoption, and measurable business outcomes. Watch credit conditions for signs financing risk is rising, and treat trust, safety, and local acceptance as operating constraints - not side issues.
If you're making AI decisions inside your own business, the next step is simpler and more useful than a broad adoption target: pick one costly workflow, establish a baseline, measure full cost, and set a clear decision rule. Expand only when the evidence supports it.