Hyperscalers Just Committed 700 Billion Dollars. Only 64 Percent of Buyers Track Their Own AI Spend.

Week of 3 August 2026 | Compiled by Team Paringa

Earnings week settled the supply question and reopened the pricing one. Capacity is being built at unprecedented scale, and someone has to pay for it.

SECTION 1 - What Happened This Week

🔴 Earnings week: record capex, accelerating cloud growth, and a market that stopped rewarding everyone

Between 28 and 31 July the four largest hyperscalers reported and the numbers reset every cloud budget assumption for 2027.

Amazon reported on 30 July. AWS revenue reached 42.23 billion dollars in the June quarter, growing almost 37 percent, the fastest expansion since 2021 and a sharp acceleration from 28 percent in Q1. Operating income was 16.62 billion dollars against a 13.62 billion consensus, at a 36.8 percent operating margin. AWS backlog hit 496 billion dollars. The AI business and Amazon's own silicon each passed 25 billion dollars in annualised revenue, more than doubling year on year. Amazon then raised full-year 2026 cash capex guidance to approximately 220 billion dollars, up from the 200 billion previously signalled.

Microsoft reported Q4 FY2026 on 29 July: revenue of 90.0 billion dollars, up 18 percent, operating income 40.6 billion, up 18 percent, and net income 35.8 billion, up 31 percent. Azure grew 43 percent in constant currency against a 40.3 percent consensus, and CFO Amy Hood guided to 45 percent for the current quarter. Capital spending hit 41 billion dollars in the quarter, up 70 percent year on year and 28 percent above the 31.9 billion reported the previous quarter. Roughly two-thirds of that was short-lived assets, meaning CPUs and GPUs, with about 5.6 billion in finance leases.

The market reaction is the part CFOs should read closely. Alphabet reported first, on 22 July, and its 195 to 205 billion dollar 2026 capex estimate triggered an immediate sell-off that set a sceptical tone for the rest of the week. Meta raised full-year capex guidance to 130 to 145 billion dollars and fell 8 percent. Microsoft rose 15 percent. Alphabet, Amazon and Microsoft added close to 1.5 trillion dollars in combined value across the week. Investors are no longer paying for capex as a proxy for AI ambition. They are paying for capex that comes attached to demonstrable revenue.

Combined 2026 AI capex across Amazon, Alphabet, Meta and Microsoft is now tracking around 725 billion dollars, up roughly 77 percent from about 410 billion in 2025.

Sources: CNBC - AWS Q2 2026, CNBC - Microsoft Q4 2026, The Register, CNBC - capex scrutiny

Read the asset mix, not the headline number. Two-thirds of Microsoft's capex is short-lived GPU and CPU hardware, which depreciates fast and has to be recovered through pricing inside a three to five year window. That is the mechanism by which 725 billion dollars of infrastructure becomes your renewal quote. Amazon's 496 billion dollar backlog tells you commitment discounts are being used aggressively to lock in demand ahead of that recovery, which means the negotiating leverage available to you now is better than it will be in eighteen months. If you have a renewal or a commitment decision inside the next two quarters, model it against a scenario where unit pricing rises rather than falls. More supply is not the same as cheaper supply when the supply was this expensive to build.

🟡 EY: 82 percent of leaders worry about token costs, 36 percent have no guardrails

EY published the fifth wave of its US AI Pulse Survey on 28 July, covering 534 senior decision-makers at SVP level and above across health, life sciences, energy, TMT, public sector, retail, manufacturing, financial services, private equity and real estate. The findings describe a discipline gap, not an appetite gap.

Of leaders whose organisation is investing in AI, 82 percent say the organisation is concerned about token usage and related costs. Among those using token-consuming tools, 98 percent say those costs have caused the organisation to reconsider its approach. And yet only 64 percent report that their organisation actively monitors token usage with clear budgetary guardrails in place. That leaves more than a third of enterprises consuming a metered, variable-cost resource with no instrumentation.

The direction of travel is not retreat. Among leaders whose organisations use token-based tools, 37 percent are expanding the scope of their AI rollout because of cost visibility versus 15 percent reducing it, and 29 percent are accelerating versus 15 percent slowing down. But actual spend has undershot forecast badly. A year ago 35 percent expected to be spending 10 million dollars or more on AI by now; only 23 percent are. A year ago 18 percent planned to commit half their total budget to AI; only 3 percent have. Meanwhile 76 percent say off-the-shelf software no longer meets their needs and 82 percent expect per-seat SaaS pricing to become less relevant in their industry.

Sources: EY

The 82 percent concerned versus 64 percent instrumented gap is the whole FinOps opportunity in one statistic. Concern without telemetry produces the worst outcome available: programmes cancelled on gut feel rather than trimmed on evidence. The fix is unglamorous - per-team and per-use-case token attribution, a routing policy that sends the bulk of traffic to cheaper models and reserves frontier models for work that genuinely needs them, and a hard budget ceiling with alerting before the invoice arrives. Note also that 34 percent flagged shadow IT as a top barrier to in-house AI build; shadow AI spend on corporate cards is real money and it does not appear in the cloud bill. Get visibility of both before you make any decision about scope.

🟡 Oracle's power commitments show where the next cost increase comes from

On 21 July The Register reported that Oracle faces a roughly 100 million dollar annual bill to underwrite power promises tied to its Wisconsin datacentre build. That figure is worth holding next to Oracle's capex guidance, which has risen to between 90 and 95 billion dollars for fiscal 2027, which began in June, up from an earlier forecast of 60 billion.

This is the part of the AI infrastructure story that has not yet reached most cloud invoices. Energy contracts, grid connection commitments and local underwriting obligations are being signed now at multi-decade durations, and they sit underneath every GPU hour that will be sold from 2027 onwards. Analysts tracking sector-wide AI datacentre capex put the figure at around 1.5 trillion dollars over five years to 2030.

For platform and FinOps leaders the implication is straightforward. Compute unit economics over the next three years will be shaped less by chip supply and more by the cost of the power and the property that surrounds the chips, and those costs are far less elastic. Regional price variation between availability zones will widen accordingly, and workload placement becomes a lever it has not been before.

Sources: The Register - Oracle Wisconsin, The Register - AI bubble warnings

Start treating region selection as a cost lever, not a latency decision. Power pricing diverges regionally in a way that compute never did, and the gap will show up in list pricing within two years. Two practical actions this quarter: tag every workload with a residency constraint so you know what can actually move, and get a real inference cost per transaction for your top three AI use cases. Most organisations cannot produce that number, which is precisely why they cannot tell whether a use case is worth running.

SECTION 2 - What CFOs, CIOs and FinOps Leaders Are Dealing With

The cheap wins are gone. The FinOps Foundation's State of FinOps 2026, drawn from 1,192 organisations representing more than 83 billion dollars of annual cloud spend, records that practitioners have already cleared the large, obvious waste and what remains is smaller savings spread across many more workloads, requiring more analysis per pound recovered. In parallel, Flexera's 2026 State of the Cloud puts wasted IaaS and PaaS spend at 29 percent, up from 27 percent in 2025, because AI workloads make forecasting structurally harder. Seventy-two percent of organisations exceeded their cloud budgets in the last fiscal year.

The share of FinOps practitioners managing AI spend went from 31 percent in 2024 to 98 percent in 2026. That is a discipline absorbing an entirely new cost class in two years, and it is doing so while the underlying pricing model shifts from predictable per-seat subscription to variable, usage-metered consumption. The organisations coping best are the ones that put attribution in place before scale rather than after the first surprise invoice.

Three Things Worth Watching Next Week

1. Post-earnings commitment repricing - With guidance now public through FY2027, watch for hyperscaler sales teams pushing longer reserved and committed-use terms. That is where the leverage window opens and closes.

2. Alphabet and Meta capex justification - Both were punished for spending without a clear revenue link. If either revises guidance or reframes the AI revenue story, it signals how much discipline the market is going to force on the sector.

3. Inference pricing moves - Frontier model providers have been absorbing margin pressure. Any list-price change on inference or context-window pricing lands directly in enterprise AI budgets with no procurement cycle to absorb it.

Questions? hello@paringa.ai

Paringa works with enterprises on infrastructure and AI cost, security and governance through a partner ecosystem covering cloud and AI cost optimisation. We help you see what you are spending before you decide what to cut. | 3 August 2026 | © Paringa Ltd