Expensive Oil, Cheaper AI: The Battle to Explain the Bill

This Saturday, the bill matters more than the demonstration. Oil finished Friday near $117 and Wall Street lost ground again, while IBM argued that many businesses can obtain sufficiently capable AI for much less money. These are not opposing stories: both ask what it costs to turn capacity into results. The argument for October 24 is that the next phase of the digital economy will be decided as much in budgets as in models. At 7 a.m. Madrid time, with US stock exchanges closed for the weekend, Friday's known closing figures must be separated from threats and expectations that remain unresolved. The technological opportunity is still there. What is changing is the patience available to finance it without a convincing explanation of its return.

Macro/Energy: oil also raises the cost of patience

Associated Press, in its report published by BNN Bloomberg, puts Friday's Brent settlement at $116.72 a barrel, up 3.1%. The Guardian describes an intraday move above $117 followed by some moderation: that is an earlier snapshot of the session, not an alternative settlement price. Both reports agree on the direction, although their descriptions of the historical high differ; this edition does not use that superlative. The important mechanism is that more expensive energy can increase transport and production costs while leaving households less money for other purchases. That combination complicates central banking because cooling demand does not produce barrels. For a European business, watching crude alone is insufficient: exchange rates, contracts, transportation and the time taken for higher costs to reach its bill also matter. Transmission is neither automatic nor identical for everyone.

Geopolitics: a threat is not yet a disruption

According to The Guardian, Tehran warned of possible attacks against oil facilities, subsea cables, data centres and AI infrastructure in the Gulf if a new US-Israeli offensive took place. The conditional matters: the warning does not establish that those facilities have been attacked. The same report attributes to the Revolutionary Guards a claim to have struck four commercial vessels; that statement is not treated here as independently verified either. Markets can incorporate a risk premium before an additional physical supply loss occurs, and that premium can retreat without the conflict ending. Three questions therefore need separate answers: what is threatened, what is confirmed, and what physically changes in production or transportation? The business implication extends beyond oil. A region concentrating energy, connections and computing capacity can bring together dependencies that a contract presents as separate services. Diversifying suppliers does not always diversify geography.

AI/Tech: the strongest model does not always deliver the best margin

Xataka reports remarks by Arvind Krishna during IBM's earnings presentation: the executive estimates that open-weight models achieve around 90% of the effectiveness of the most advanced models at between 10% and 25% of their per-token cost. This is an assessment by a seller with interests in the market, not a technical law. The article notes that Krishna did not identify specific tests allowing the comparison to be reproduced. Open weights also do not necessarily mean open source or access to the entire training process. The significance is not that these percentages become a universal promise, but that the question is changing: for summarising documents, classifying messages or helping answer enquiries, how much quality does the user actually require? Always buying the frontier model can be as unwise as dismissing its capabilities when the task requires precisely that additional performance.

The fine print is in the denominator. A cheap token does not guarantee a cheap task if completion requires more attempts, more context or more human review. Running a model internally adds equipment, electricity, maintenance, security and people capable of operating it. Xataka reports IBM discussing enterprise deployments using a few racks, but also notes that an API can remain cheaper at low utilisation. A useful comparison would evaluate alternatives on the same real work, establishing a quality threshold beforehand and counting failures and review time. An internal draft and a contract analysis do not necessarily tolerate the same error rate. The energy connection is direct but not mechanical: a barrel's price does not by itself determine a data centre's electricity bill, yet both remind us that digital intelligence has physical and operational costs. Efficiency counts when it survives that complete accounting, rather than simply looking attractive on the provider's price list.

Markets: Friday was not a replay of Thursday

Associated Press reports Friday declines of 0.5% for the S&P 500, 1.2% for the Dow Jones and 0.2% for the Nasdaq Composite. The last index must not be confused with the Nasdaq 100. The same report records gains in Micron, Intel, Microsoft and AMD, while almost three out of four S&P 500 constituents declined. That distinction matters: improvement in selected technology stocks can cushion an index without describing a broad market recovery. Oil supported some energy companies, and the ten-year Treasury yield finished at 4.99%, compared with 4.96% on Thursday, according to that coverage. These are closing figures, not live Saturday prices. The combined signal is not that all interest in AI has disappeared. Selective buying, energy pressure and financing demands coexist. A partial rebound does not erase the risk behind the preceding sell-off either.

Charles Schwab's Friday opening commentary provides context, not another closing report: it describes concern about large technology companies' spending and the time required to recover investment. That concern fits IBM's debate without settling it. If business customers obtain useful intelligence more cheaply, they may expand adoption; some suppliers could simultaneously face pricing pressure. A cheaper service can enlarge a market without enriching every participant equally. Investors need to distinguish demand for capacity, revenue and cash generation: these are connected stages, not synonyms. Business managers face a similar consequence. A project does not improve because it produces more answers if it also multiplies corrections. The shared question for markets and operations departments is when growth becomes a sustainable result after infrastructure, financing and supervision have been paid for. That is a more demanding test than showing that a system can perform an impressive demonstration or attract a growing number of requests.

The next 24–72 hours: prepare for Monday without inventing a calendar

Over the weekend, the first watchpoint is verifiable confirmation concerning supply, navigation or infrastructure damage in the Gulf. Not every statement requires a changed plan; evidence affecting delivery times or availability would be more consequential. The second is market reopening and the reaction to risk accumulated outside stock exchange hours, without predicting its direction here. The third is the corporate calendar: Schwab lists Verizon earnings on Monday, October 26, an early examination of investment and connectivity. Its calendar also includes major technology companies and the Federal Reserve later in the week, but these fall outside the core of the next 72 hours and are not events happening today. Dates and times should be confirmed against official sources before trading. For businesses, the weekend also offers a less conspicuous task: checking which projects depend on a single route, region or platform, and how much room exists to reschedule them if conditions change.

Conclusion: three scenarios and one small decision

In the base case, energy tensions continue without an additional sustained disruption and businesses keep adopting AI with greater cost discipline. The practical implication would be to stagger commitments and require a cost per accepted task, not an activity figure. In the bullish case, verifiable de-escalation eases logistical pressure and cheaper models expand profitable uses without weakening quality. Scaling projects already demonstrating those conditions would make sense, without assuming every supplier will benefit. In the bearish case, escalation causes fresh disruptions while technology expenditure takes longer to generate cash. Contractual flexibility, operational alternatives and fewer concentrated dependencies would become more valuable. These are analytical scenarios, not calculated probabilities or recommendations on individual securities. Evidence allowing a change of scenario matters more than correctly choosing a label today for the entire quarter. Each scenario is useful only if it connects a development that can be observed with a response that can actually be implemented.

Saturday's practical application fits on one page: choose a repetitive task, write down the result it must deliver, and calculate the cost of achieving the required quality, including the person reviewing it. Then write what would happen if the supplier were unavailable for a working day. That exercise turns an abstract discussion about oil, geopolitics and open models into a manageable decision. Predicting the next headline is unnecessary to identify spending that adds no value or a dependency nobody has checked. Prudence should not be confused with immobility either: experimenting on a small scale can be the cheapest way to discover where an advantage genuinely exists. Useful composure means reserving attention for what can be verified and commitments for what already demonstrates results. This weekend's underlying story is not that technology promises less, but that explaining its bill is beginning to matter as much as demonstrating what it can do.

Sources and scope

Editorial cutoff: October 24, 2026, 7 a.m., Europe/Madrid. Friday closes are from AP; geopolitical statements are attributed to their source and IBM percentages are not an independent benchmark. Scenarios and implications are editorial analysis. Current coverage was selected through web search and the Xataka feed.