TheChek News Flow

DAX:

MDAX:

SDAX:

GRAHAM ADVISORS | EQUITY RESEARCH | 2 OCT 2026

NKE US: GUIDE CUT 25%, STOCK DOWN 9%. A BOUNCE FROM $30, A DECISION IN NOVEMBER

================================================================================

Close $35.15 | after hours $32.10 (-8.7%) | lowest since 2013 [T2,T3]

BOTTOM LINE

No rush to own Nike. At $32 the stock pays about 26x a guided trough, the

hardest quarter is still ahead, and only the Investor Day on 16-17 Nov can

re-anchor normalised earnings. A bounce is tradable, but from $30 rather than

$32. In the sector the de-rating is broad, with one name left out: JD Sports

(JD/ LN) takes over 40% of its sales from Nike and is down 7% this year,

against 44% for Nike. That gap is the relative trade.

================================================================================

MAIN POINTS

================================================================================

1. THE GUIDE IS THE STORY, AND THE BEAT WAS THIN. Q1 EPS $0.48 vs $0.44, but

operating profit fell 4% and pre-tax income was flat only because other

income swung by $42m. FY27: revenue guided down high single digits,

adjusted EPS $1.15-1.35 vs $1.67 consensus. That is 25% below at the

midpoint, 40% below FY26, and before about $0.15 of Pace charges. [T1,T2]

2. THE GUIDE IMPLIES AN AIR POCKET. Q1 revenue fell 4%. To finish the year

down 7-9%, the next nine months must fall 8-11%, which removes about $3.7bn

of sales in a year. The last three quarters add up to $0.67-0.87 of EPS,

and Q2 carries an extra 400bp headwind. [GA est. on T1]

3. NORTH AMERICA IS NEXT. It grew 2% only on sell-in: wholesale +9%, Nike's

own stores and digital -6%. Order books are lower, Jordan retro launches

are being cut, and China is guided worse than its 22% decline. On our

arithmetic the guide midpoint leaves North America down about 4-7% for the

rest of the year. [T1, GA est.]

4. "DELIBERATE" EXPLAINS ONLY PART OF IT. The Dunk cut was a $200m headwind, a

quarter to a third of the Sportswear decline (just under half of sales,

down low double digits). The rest is product and aged inventory: Dick's has

said classic-colour Air Force 1s were sluggish last quarter. Inventory is

down 3% while sales head down 8-11%. [T1,T3, GA est.]

5. MARGIN, AND WHAT PACE IS WORTH. The guide implies an EBIT margin of about

5.7%, against 8.3% in FY26 and 12.7% in FY24: a gap of about $3bn a year.

Pace is $2.5bn "cumulative" through FY31, before $1.0bn of charges and

before reinvestment, mostly in FY29-30; job decisions only start in

calendar 2027. As a five-year sum that is about 1 point of margin; as a

run-rate it is close to 6. The release does not say which: first question

for November. [T1,T3, GA est.]

================================================================================

THE STOCK

================================================================================

- THE MULTIPLE WENT UP. 21x the old consensus at the close, about 26x the

guided midpoint at $32.10. On a 10% margin (about $2.25 of EPS) it is 14x,

and that is an FY29-30 number on management's own timetable. [GA est.]

- THE YIELD IS NOT A FLOOR. $1.64 is a 5.1% yield with the 10-year Treasury at

5.25%, and a 131% payout: about $2.4bn a year against about $1.9bn of guided

net income. Cash is $8.4bn but debt is $7.9bn, with $2.0bn due within a

year; cash fell $0.66bn in the quarter. The CFO says the dividend is

supported in every scenario. [T1,T2]

================================================================================

$30: A LEVEL TO TRADE, NOT YET A FLOOR

================================================================================

- FOR. At $30 the yield is 5.5%, EV falls below 1.0x trailing sales, and the

stock is on 13x a 10% margin, close to the 10-12x the market pays for peers

that earn their margins today. The price chart adds the round number and the

2013 base. [GA est.]

- AGAINST. Nothing has traded there since 2013, each of the last three

post-results lows was broken within months, and FY28 estimates still have to

come down. [T2]

- THE ARITHMETIC. From $30, with a stop under $28.50 and the gap at $35 as

target, the bounce pays about 3 for 1. From $32.10 it pays less than 1 for

1. [GA est.]

- TRIGGERS AND EXIT. After the last two guide-downs the lead director and the

CEO bought within two weeks, at $59-61 in December and at $42 in April. The

stock gained 5-7% each time, then gave it back. Next: the run-up to 16-17

Nov and the dividend decision (announced on 20 Nov last year). The trade

ends before the December print. [T2,T3]

================================================================================

SECTOR: A BROAD DE-RATING, WITH ONE NAME LEFT OUT

================================================================================

Since 2 Jan, in dollars:

LULU -55% | NKE -44% | ONON -36% | DKS -34% | DECK -25% | ADS -16% | JD -7% [T2]

Results sold since late July, good ones included: ADS -11.5% on record sales,

DECK -6% on a beat and raise, ONON -20% on a trimmed outlook, JD -14% and LULU

-17% on guidance cuts. [T2,T3]

Multiples on guided earnings: LULU 10x, DECK 11x, DKS 12x, JD 7x, on margins

they earn today. NKE 26x, on a margin it hopes to recover. [GA est.]

READ-ACROSS

(-) NIKE-DEPENDENT US RETAIL: JD/ LN (Nike over 40% of sales, North America

like-for-like -6.8%), Foot Locker within DKS US (guided to a loss). Jordan

launches cut, sell-in turning negative, aged sportswear to clear.

(-) CHINA PARTNERS: Topsports 6110 HK (22% of revenue), Pou Sheng 3813 HK

(15%). Greater China -26% currency neutral, wholesale -31%; partner online

stores closed from 1 Jan 2027.

(-) SUPPLIERS: Feng Tay 9910 TT, Yue Yuen 551 HK (H1 profit -58%). $3.7bn less

Nike revenue; a "more flexible" supply chain under Pace.

(+) SHARE TAKERS (RELATIVE): ADS GY, Chinese local brands by inference. The

China fall is Nike-specific: Adidas China +15% in Q2, Performance +39%.

(?) RUNNING CHALLENGERS (WATCH): DECK US, ONON US. Nike's share of max

cushioning "nearly tripled"; HOKA slowed to +8%, On Americas to +13%.

Precedent for the open: on Nike's April guide-down (NKE -15.5%), JD -0.2%, DKS

-3.1%, ADS +0.2%. [T2]

================================================================================

THE PAIR: LONG NKE US / SHORT JD/ LN (CHART ATTACHED)

================================================================================

- THE GAP. Since 2 Jan Nike has lost 44% and JD Sports 7% in dollars. The pair

is down 40%, near its low for the year, and at a new low at Nike's

after-hours price. [T2]

- A PAIR, NOT A HEDGE. JD fell 0.2% on Nike's 15.5% drop in April, and its

weekly correlation with Nike is 0.2. No listed peer hedges Nike this year:

On is the closest match (correlation 0.5, beta 0.6), and it rose on the

April guide-down. Nike's own risk is covered only by the stop or by puts.

[T2, GA est.]

- THE CASE IS CATCH-DOWN. Nike is over 40% of JD's sales and North America

38%, where like-for-like fell 6.8% last quarter. Nike is cutting Jordan

launches and order books, and its guide points to sell-in turning negative.

JD's GBP700-800m guidance leaves 60-65% of the year's profit to the second

half, and the stock rallied 9% in the six sessions before Nike's print.

[T1,T3, GA est.]

- THE RISKS. JD trades on about 7x earnings, with net cash and a GBP200m

buyback, and the pair has lost money since April. JD updates on 19 Nov, two

days after Nike's Investor Day. [T1, GA est.]

HOW EACH CANDIDATE SHORT HAS TRACKED NIKE IN 2026

SHORT WEEKLY CORR BETA TO NKE 1 APR MOVE PAIR SINCE 2 JAN

-------- ----------- ----------- ---------- ----------------

JD/ LN 0.22 0.24 -0.2% -40%

DKS US 0.25 0.36 -3.1% -16%

ONON US 0.47 0.62 +2.2% -14%

DECK US 0.29 0.37 +0.8% -26%

LULU US 0.26 0.34 +3.7% +22%

ADS GY 0.36 0.36 +0.2% -33%

38 weekly returns to 25 Sep. 1 Apr = session after Nike's Q3 guide-down (NKE

-15.5%). Pair = long NKE, short the name, in dollars; ADS to 30 Sep. [T2, GA

est.]

================================================================================

ASYMMETRY, 12 MONTHS (GA SCENARIOS)

================================================================================

SCENARIO PROB VALUE VS $32.10 VS $30 ASSUMPTION

-------- ---- ----- --------- ------ ------------------------------

BEAR 30% $22 -31% -27% Reset runs through FY28, EPS

about $1.15, dividend in question

BASE 45% $35 +9% +17% Guide delivered, FY28 EPS

about $1.65 at 21x

BULL 25% $44 +37% +47% Credible path to a 10% margin,

about 20x $2.25 paid early

Weighted value about $33: +4% from $32.10, +11% from $30. Upside to downside

moves from 1.2x to 1.8x.

================================================================================

WHAT WOULD MAKE US WRONG

================================================================================

- The reset is taken as the clearing event and the stock squeezes from $32

without touching $30.

- $30 breaks and Nike converges to the sector multiple: 10-12x normalised EPS

is $22-27.

- FY28 is guided down again in November and the dividend becomes the debate.

- Holiday sell-through of the running and basketball launches surprises on the

upside.

- JD keeps decoupling: the market goes on treating Nike's reset as Nike's

problem alone, as it has since April.

================================================================================

T1 company filings and calls | T2 market data and consensus | T3 press

GA est. = Graham Advisors estimates

LC / Graham Advisors

GRAHAM ADVISORS | INSTITUTIONAL RESEARCH

CONFIDENTIAL | NOT INVESTMENT ADVICE

The startup, in the throes of responding to rogue AI incidents, said three employees violated policies for handling sensitive company data

OpenAI has fired three researchers for alleged misconduct, including sharing confidential company information with a third-party AI-safety organization, according to people familiar with the matter.

The company recently told some employees that it had terminated three researchers who worked on its safety team, one of the people said.

The affected employees are Jasmine Wang, Tomek Korbak, and Mikita Balesni, the people familiar with the matter said. The researchers didn’t immediately comment.

“We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” an OpenAI spokesperson said. “Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.”

AI giants are under growing pressure to submit their technology for independent safety testing.

Last month, the CEO of OpenAI rival Anthropic said his company would allow outside evaluators like METR, an AI safety nonprofit, to verify its adherence to safety measures and assess model alignment.

OpenAI has recently faced a spate of security incidents in which its AI agents escaped containment, hacking some company websites and aggressively probing a wide range of websites.

After its model hacked the AI company Hugging Face, OpenAI allowed staff members from METR and a Redwood Research staff member contracting with that group to work in its offices for six days to investigate how its models behaved. METR later released a report based on the information to which the company provided access.

One of the fired OpenAI employees, Korbak, was a member of OpenAI’s safety team, and has said he served as the company’s technical contact for Redwood Research and METR in their investigation of the Hugging Face incident.

The other terminated employees, Wang and Balesni, worked on alignment, which is ensuring that the company’s models behave as humans intend them to.

OpenAI has said that it is working to investigate a range of agent security incidents that it has discovered in recent months, and address underlying safety issues. Earlier this week, OpenAI said it was scrapping the planned launch of an AI model, GPT-6.1 Astra, over safety concerns.

The Chat GPT-maker said it has implemented a new monitoring system to catch AI-agent misbehavior more quickly, started requiring engineers to use stronger security guardrails for testing its AI systems and is sharing more information about instances in which models behave badly.

The AI industry is grappling with the growing capabilities of its most powerful models and the risks they pose.

In early September, Anthropic researcher Jacob Coxon publicly quit, saying he didn’t want to participate in a rush to build AI systems that can improve themselves. He said he feared they could spiral out of control and destroy humanity.

Anthropic Chief Executive Dario Amodei wrote last month that the risks posed by today’s cutting edge AI tools were too great to continue development at the same breakneck pace and called for slowing down industrywide development, drawing agreement from OpenAI Chief Executive Sam Altman and Elon Musk.

Security experts say an assailant with a plan and a knife, as alleged in Wednesday’s flight to Israel, shouldn’t have made it as far as the airport

The FlyDubai pilot who stabbed his captain and sent the plane diving toward the ground slipped through one of the world’s tightest aviation-security nets. The question is: How?

The route connecting the United Arab Emirates and Israel is one of the world’s most sensitive. The Middle East has a long history of hijackings and other attacks on civilian airliners.

To counter attacks, both countries have some of the world’s most advanced surveillance technologies and capable spy networks. On top of that, air regulations have been repeatedly tightened following the Sept. 11, 2001, terrorist attacks on the U.S. and the Germanwings Flight 9525 intentional crash in the French Alps.

And yet on Wednesday a pilot was able to conceal his intentions, gain access to a knife, and put a plane carrying around 170 people into a dive so steep it nearly tore it apart in midair.

“This is 100% a major security failure,” said Michael Horowitz, a security and geopolitical analyst who wrote a book about Israel’s future in the Middle East. “The question is how broad and how bad.”

The answer will weigh on Israel’s air links to the world and the prospects of the U.A.E.’s core aviation and tourism businesses. FlyDubai suspended its flights to Israel while the investigation continued. A number of international airlines already had suspended flights to both the U.A.E. and Israel for months because of the risks associated with the Iran war.

Saudi officials initially assessed that the co-pilot tried to deliberately crash the plane after stabbing the pilot, people familiar with the matter said. They said they didn’t yet have a theory about why. The U.A.E. said Thursday it was investigating whether terrorism or some other motive played a role in the attack.

If it does turn out to have been a terrorist act, “this would be the most significant terrorist plot involving a commercial passenger aircraft in recent years,” aviation security firm Osprey Flight Solutions said in a report.

FlyDubai didn’t respond to requests for comment.

Aviation experts consider Dubai International Airport to have high security standards. It was consistently the world’s busiest for international travel and connected fliers from risky countries such as Afghanistan and Iran. It uses facial recognition to identify passengers and scans carry-on luggage coming in and out of the country.

Pilots who have regularly flown from Dubai recounted strict searches through metal detectors that have led security staff to confiscate everything from selfie sticks to hair spray to cups of coffee. Air regulations ban all blades with sharp points and typically blunt knives longer than six centimeters.

“Dubai is high-spec and a key global hub. It compares well to many of the top airports in the world,” Horowitz said. “But that doesn’t rule out insider jobs or a broader lapse in security.”

An Israeli official said the country has long pressed Dubai airport officials to strengthen security for flights heading its way. It stopped its own airlines from flying between Dubai and Tel Aviv for several months this year amid worries about security gaps and the risks posed by the war against Iran.

Israeli airlines typically follow stricter security protocols than foreign airlines flying the same routes.

Starting Friday, Israeli airlines will begin operating limited rescue flights to bring out Israelis stuck in Dubai after FlyDubai suspended service, according to Israel’s Ministry of Transportation. Israel’s flagship airline, El Al, said it wouldn’t allow passengers to bring checked or carry-on luggage at the instruction of security officials.

Israel has also taken issue with security on foreign airlines. In May 2024, Israel’s comptroller published a classified report that found major problems with the security of flights to and from the country, including inadequate security at foreign airports, a person familiar with the matter said. It found that Israel’s Ministry of Transportation was focused on security on Israeli airlines and wasn’t investing enough to secure foreign airlines that carried Israeli passengers. The Israeli prime minister’s office didn’t return a request for comment on the report.

Dubai’s media office and airport didn’t respond to requests for comment.

The Sept. 11, 2001, attacks in which hijackers seized control of jets and flew them into the World Trade Center and the Pentagon reshaped the rules around international air travel. A key change was the introduction of reinforced doors to keep intruders out of the cockpit.

The regulations mostly revolved around protecting aircraft from passengers, however, not pilots or crew who might present threats from within.

The 2015 Germanwings crash, in which a first officer took control of the cockpit after being left alone and intentionally flew the plane into the French Alps, killing everyone on board, led some regulators to require that two crew members be in the cockpit at all times. The pilot who crashed the plane had received psychotherapy for suicidal tendencies, and the industry has grappled with mental-health risks ever since.

Pilots typically are required to undergo a medical screening at least once a year but face suspension of their licenses if they receive certain mental-health diagnoses or are prescribed related medication. That has prompted concerns that aviators could hide symptoms and avoid seeking help.

“We don’t know if it was a terrorist plot, if he intended to crash the plane, if it was just an altercation,” said Andy Blackwell, an analyst at aviation-security specialist ISARR and a former head of corporate security at Virgin Atlantic. “We don’t know their stresses. We don’t know their state of mind.”

FlyDubai has an in-house psychology division that vets first officers as they graduate to captain, according to a former pilot at the company and an active crew member at the sister carrier, Emirates. It wasn’t immediately clear whether the assailant would have undergone such a review.

A central question in the investigation is how the assailant was able to get the knife aboard. In addition to screening, airports take inventory of knives in restaurants located after security, and careful decisions guide what planes may carry.

It is possible the assailant smuggled his own knife on board or used a blade that was already post-security and accessible to the crew. He also could have had help.

“Not everything that needs to be detected is detected,” said Matthew Borie, chief intelligence officer and co-founder of Osprey Flight Solutions. “Insiders typically know how to breach the measures that are in place, because they’re part of the system.”

Screening is the last line in the security funnel. If a bad actor made it that far, there was already an intelligence failure in allowing a plot to go forward, security analysts said.

Dubai keeps close watch on its residents. When a Hamas militant turned up dead in a Dubai hotel room in early 2010, authorities combed through security-camera footage and electronic-payment receipts to quickly assemble images of the more than two dozen members of the hit team, even some that showed them donning disguises. The assailants had escaped, but it was a strong show of the city’s capabilities.

Borie said a significant plot shouldn’t have made it to the airport without being exposed.

“If this was really as what has been stated by the Israelis as a jihadist terror attack, the idea that this needed to be detected solely by the last level of screening before somebody boarded the aircraft—that’s the last layer of defense,” he said. “When insiders are involved, we’re putting an extensive amount of pressure on the lowest rung of the ladder to make the stop.”

Supply chains that transform crude into fuel are still under siege, and that is keeping prices high—helping drive a brutal bond selloff

Global crude futures jumped 4.4% to $102.31 a barrel on Thursday as supply-chain disruptions keep fuel prices high.

The 10-year Treasury yield notched its highest level in 24 years this week, driven by persistent inflationary pressures from the oil market.

Colliding threats include Houthi rebel attacks on shipping lanes and a Ukrainian air campaign targeting Russian refineries.

The return of oil tankers through the Strait of Hormuz was supposed to bring down the price of crude, gasoline and diesel, taming inflation and lowering the world’s borrowing costs. It isn’t happening.

Even as U.S. forces have worn down Iran’s grip on the waterway, helping oil shipments rebound toward prewar levels, the global crude-futures benchmark jumped 4.4% Thursday to $102.31 a barrel. Average U.S. gas prices tracked by AAA were hanging above $4.41 a gallon. And the 10-year Treasury yield, a rough proxy for interest rates on many types of loans, this week notched its highest level in 24 years.

The factors helping keep bond yields sticky stem from interlocking threats that are still squeezing the oil market like a vise:

• Estimates for Hormuz traffic vary, and traders fear re-escalation could throttle back supplies at any moment.
• Houthi rebels are threatening shipping lanes that also carry oil across the Arabian Peninsula.
• Thousands of miles away, a Ukrainian air campaign is targeting Russian refineries, which have historically been diesel factories feeding many industrial economies.

Those colliding factors have pushed up the cost of moving crude and making fuel across the supply chain, raising the price to operate everything from cars and trucks to commercial jets and farming equipment.

The bottleneck is sending inflationary aftershocks across America that could extend beyond November’s midterm elections, pressuring the Federal Reserve to consider additional interest-rate hikes in the months ahead.

Already, the bond market is struggling to absorb a flood of artificial-intelligence-linked debt, huge fiscal deficits and an economy hot enough to keep inflation above the central bank’s target. While the 10-year Treasury yield Thursday ticked lower to 5.233%, it is fresh off its steepest one-quarter run-up since 1994, helping propel U.S. mortgage rates past 7%.

“There is still a high degree of sensitivity in the markets now to whether there’s progress to some kind of formal agreement between the U.S. and Iran, to the potential for further disruption,” said Arend Kapteyn, global head of economics and strategy at UBS Investment Bank.

“From a central bank perspective,” he added, “you don’t just care about crude, but you care about the (refined) products, and the products are super tight.”

As the U.S. Navy has degraded Iran’s ability to attack tankers or nearby energy facilities, Gulf oil producers and shipping firms have grown more adept at fending off or evading strikes. Morgan Stanley analysts estimate Middle East crude exports have recently been just 7% below prewar levels.

But with many tankers turning off tracking systems or transferring cargoes between ships to minimize threats, other analysts are circumspect about just how much shipping has recovered. The flow data “could be a little optimistic, but it’s also not necessarily going to be consistent,” said Rebecca Babin, a senior energy trader at CIBC Private Wealth. “It’s really opaque.”

Renewed strikes on regional shipping lanes by Iran or its Houthi proxies in Yemen are also pushing traders to bake in more risk to their positions—buoying prices. On Wednesday, the U.K. Maritime Trade Operations Centre reported three vessels in the Strait of Hormuz were hit with projectiles.

Drawdowns in inventories from China to Europe to the U.S., which buffered the market in the conflict’s first seven months, have left little room for error. Those stockpiles will also need to be refilled.

The uncertainty has left traders like Babin glued to headlines about possible fuel-export restrictions by the U.S. and China. Even though crude futures have retreated somewhat from recent highs, the cost of physical oil cargoes remains more elevated, signaling refiners are racing to lock in short-term supplies. North Sea dated crude—a benchmark reported by Argus Media to track on-the-spot deliveries—fetched $127.42 a barrel Thursday.

“There’s clearly still this demand for these barrels regardless of what the numbers say is coming through [Hormuz],” Babin said.

While tanker traffic has rebounded, J.P. Morgan recently told clients, “The recovery, however, is uneven.” The bank estimates that exports of refined products such as gasoline, diesel and jet fuel from the region still remain around 40% below prewar levels.

The cost of bringing crude to market has also skyrocketed. Tankers heading west from the Gulf have at times taken a long trip past the southern tip of Africa to avoid the threat of Houthi attacks. At the same time, many shipping firms are carrying oil out of the Gulf by shuttling supplies just outside Hormuz, where it can be transferred to vessels headed elsewhere.

Those factors have tied up much of the global fleet of supertankers and sent freight rates surging. The roughly 21-day trip by skyscraper-size vessels ferrying crude from the Middle East to China recently cost the equivalent of $35 a barrel, according to Argus, up from less than $7 the day before the war began.

Once that crude reaches refiners, costs are snowballing further. Thanks to reduced capacity in the Middle East, as well as Ukrainian strikes on Russian refineries that pushed Moscow to curb diesel exports, traders are paying a premium to companies elsewhere that can pump out fuel.

In New York, diesel futures have recently fetched roughly double the price of crude, with spreads far exceeding any previously recorded levels, boosting wholesale costs for gas stations.

Many energy executives expect the dynamic will persist; 48% believe at least a year will pass before the spread between diesel prices and crude costs normalizes to 2025 levels, according to a Dallas Fed survey released this week. Roughly 36% said the same about gasoline.

Now, some companies’ plans to hedge against higher costs that they locked in early this year or after the war’s outset are ending, said Charlie Macnamara, head of U.S. Bank’s commodities team. Businesses such as farmers, home-improvement stores and truckers face a tough choice.

“Do I lock in now? Or do I wait and hope that things get better?” Macnamara said. Higher prices now “will start to bleed into those hedging plans, and that will ultimately bleed into the end consumers.”

President Trump has said he expects Tehran won’t make a deal until after the midterms. Privately, the president also told aides he expects to resume bombing after Americans vote, The Wall Street Journal reported.

While energy prices remain painfully high, analysts believe they are not high enough to force a major slowdown in an economy fueled by the AI boom and a record-breaking stock market.

“If economic resilience and investment demand remain intact, there may still be scope for yields to move higher,” said Seema Shah, chief global strategist at Principal Asset Management.

the economics of computing

Powerful AI tools are burning through budgets, prompting a rethink of how the technology is priced ahead of frontier lab IPOs

Barely four months into 2026, Uber had already hit its annual AI spending limit.

“We budget based on what we know, but things are changing fast in this space,” said Praveen Neppalli Naga, Uber’s chief technology officer. “Every week there’s a tectonic shift.”

Uber was not alone. As companies raced to adopt AI agents, bosses introduced leaderboards and targets to encourage staff to use them. Their ability to carry out complex tasks with limited human input is “the real unlock for AI”, said Ara Kharazian, chief economist at fintech company Ramp.

That autonomy has made AI costs harder to predict. Agents can burn through tokens — the units of information a model processes — with little human oversight. One Amazon project ran 860 per cent over budget after incomplete tasks worked in the background for five months, costing more than $1mn, according to multiple people familiar with the matter.

By shifting some enterprise services from fixed subscriptions to uncapped usage-based pricing earlier this year, leading AI labs pushed the risk of longer workflow costs on to customers. Early adopters including Uber, Meta, Cisco and Walmart have responded with spending caps. Intensifying competition between labs has also brought down the price of newer models.

The bigger problem is how AI is priced in the first place. Most charges reflect the computing power consumed rather than the value created, prompting customers and providers to seek alternatives. Linking prices more closely to outcomes while securing durable revenues will be critical for the industry — and investors weighing anticipated initial public offerings from Anthropic and OpenAI.

“Pricing in tokens doesn’t make any sense,” OpenAI president Greg Brockman said at the September launch of its Astra model. “Price per task is what matters. Can you get the thing done at an appropriate price and appropriate speed?”

Anthropic’s long-awaited IPO prospectus highlighted the challenging economics of building state-of-the-art AI models – alongside delivering warnings about the “existential risks” they pose. The start-up recorded an operating loss of more than $8bn last year as it increased spending on computing power and told investors its rapid expansion would require $518bn in infrastructure commitments over the coming years.

At the same time, growing demand for agentic AI, particularly coding tools, has driven up revenues at the leading labs in recent months. Yet, even as AI systems become more powerful, taking on increasingly complex tasks autonomously, every interaction with a large language model starts with the same basic process.

But this process cannot expand indefinitely. The amount of information a model can handle at once is constrained by its context window — effectively its working memory.

The GPT-3.5 model that sparked the ChatGPT revolution in 2022 initially had a maximum context length of 4,096 tokens, roughly equivalent to a 3,000-word English-language essay or 20 minutes of spoken conversation.

Since then, context windows have expanded dramatically, allowing models to process far more information and enabling agents to tackle more demanding tasks. This has led to wider adoption, more token-intensive agentic workloads and greater use of more capable — and often pricier — models. Technological advances have drastically reduced the computational cost of each token, with the price for achieving the same level of AI performance falling roughly 13-fold a year since 2023 according to research group Epoch AI. But surging usage has outweighed those savings, racking up costs for companies.

Token use varies by task and model, but higher volumes generally mean greater demand for specialist chips, data centre capacity and electricity. That makes tokens a useful proxy for AI use — and a basic accounting unit of the AI economy.

“The best comparison to a token is a kilowatt-hour for electricity,” said Aleh Tsyvinski, a Yale economics professor who in July published a paper analysing the consumption habits of users who collectively consumed more than 380tn tokens across 400 different models. “It is fundamental to how we think about and measure AI’s development . . . this is the basic building block of AI.”

Companies use tokenisers — software that splits raw text into tokens — to train and run their models. The same sentence may produce a different token count depending on the tokeniser used. In every interaction with an AI system, tokens serve different functions and are priced differently depending on how efficiently they can be processed.

Cached tokens may be cheaper, but the sheer volume of usage is still pushing up costs for businesses adopting the technology. There are also other ways to manage long conversations. Systems can summarise previous exchanges, discard material not required or retrieve previously stored information, reducing the context brought forward. But summaries, retrieved material and new outputs still consume tokens — and can build up over subsequent steps.

Unlike a chatbot that returns a simple answer, coding agents such as Anthropic’s Claude Code or OpenAI’s Codex can also plan tasks, inspect files, write and revise code, and check their own work. A single request can therefore trigger far more model activity — and consume far more tokens — than a conventional prompt-and-response exchange.

OpenAI’s enterprise data illustrates the scale of the shift. As of June, Codex was used by just 17 per cent of customers, yet generated 63 per cent of all output tokens across ChatGPT and Codex.

A study involving researchers from Stanford University, Massachusetts Institute of Technology, Google DeepMind and Microsoft found that coding agents consumed 1,000 times as many tokens as simpler coding interactions.

Much of this extra activity occurs before a final answer is produced, as models repeatedly process context accumulated during a task. Yale’s Tsyvinski said this new era of agentic capabilities was driving “exponential growth” in token consumption as individuals and companies look for increases in productivity. “Agents are not writing out a basic recipe prompted by you or me, they are collecting and analysing vast amounts of data.”

This iterative cycle has become a key driver of rising corporate bills. Models do not always take the shortest route, often trying several approaches before reaching a conclusion. “You can see a meandering path of how the models get to where they have found the answer,” said Aakanksha Chowdhery, a Stanford professor and former Google DeepMind researcher. “It feels magical, but it’s kind of amusing.”

Goldman Sachs forecast in May that agentic AI would drive a 24-fold increase in global token consumption by 2030, with enterprise applications accounting for the largest share.

So far agents have seen the greatest uptake among software developers, but they will need to prove useful to other parts of a business if vast investments in AI infrastructure are to pay off. “One big driver of how quickly agentic AI disseminates through the economy will be how quickly the non-technical parts of a firm adopt agentic workflows for their tasks,” said Ronnie Chatterji, OpenAI’s chief economist.

At OpenAI, agent use spread from research and engineering to legal, finance, HR and marketing in about five months. “We’re almost all now using it,” Chatterji said, though whether other companies adopt the technology as quickly “remains to be seen”.

The rate of AI spending is rising at a faster rate among the heaviest users, according to Ramp’s analysis of 70,000 businesses. Since the start of the year, monthly spend per employee has more than doubled among the top 1 per cent of companies, to an average of $7,200, compared with a doubling across its platform as a whole.

More difficult for companies is the unpredictability of those costs. The study on agentic token consumption found that token use could vary sharply. Across repeated attempts of the same task, the most intensive run typically consumed twice as many tokens as the least, and in some cases 30 times as many. Human estimates of task difficulty were a poor guide, while frontier models also tended to underestimate how many tokens they would need.

AI agents were “not trained to be budget-aware”, said Jiaxin Pei, assistant professor at the University of Texas at Austin and one of the paper’s authors.

Taming costs

Despite several high-profile incidents — and a bout of “tokenmaxxing” in which staff faked activity to climb internal leaderboards and boost usage metrics — Amazon has pressed ahead with its enthusiastic AI rollout.

But Uber is among a number of companies that have imposed spending limits. The ride-hailing group has introduced token budgets, capping monthly spending at about $1,500 per AI coding tool.

AI usage has increased ninefold since February, the company told investors in August. In response, Uber has sought to cut costs by using services such as OpenRouter to switch between models and choose cheaper, more efficient options for different tasks. It is also caching data and asking engineers to write more focused prompts and clearer instructions — including where to find files — to avoid wasting tokens on routine work.

“Our spending has stabilised,” Neppalli Naga said. “In April [rising usage] would have been a big problem for us. We look at budgeting problems as engineering problems.”

Uber is also deploying “open” models, which can be customised and run at a fraction of the cost of the more sophisticated “closed” models from OpenAI and Anthropic. As their performance narrows the gap with leading closed models, these cheaper systems are gaining market share. Some of the most efficient have been developed by Chinese AI labs including Alibaba, DeepSeek, Moonshot AI and Zhipu AI, also known as Z.ai, drawing increasing scrutiny from regulators.

Airbnb chief Brian Chesky last year said that the company preferred to use Alibaba’s “fast and cheap” Qwen model for certain workloads, triggering a wider conversation about the defensibility of leading US models.

George Hu, president of Nvidia-backed AI infrastructure start-up Fireworks, compared using the most powerful AI for every query to asking a company’s highest-paid employee to perform every task. “The idea that every task should be done by a super expensive model doesn’t match anything we understand in the real world,” he said.

Smaller models fine-tuned for industries such as law and finance — or even individual companies — could make AI cheaper to run. Trained on specialist or proprietary data, they may require less computing power while outperforming general-purpose frontier models on narrow tasks.

As the top AI companies prepare for multitrillion-dollar IPOs and compete for enterprise customers, they are beginning to reduce costs for their top-tier models. Claude Opus 5.5 costs 40 per cent less to use than its predecessor, while OpenAI’s Sol and Luna GPT-6 versions are 50 per cent cheaper than their previous iterations.

The breakneck speed of model development, combined with usage-based billing, makes future revenues very hard to predict for Anthropic and OpenAI. “At this point, if you are looking at 2027, no one knows what they’re making,” said Joey Brookhart, analyst covering AI labs at research firm SemiAnalysis.

As well as competition from cheaper Chinese open models, a strong new release from a US hyperscaler such as Google or Meta that threatens the dominance of Anthropic and OpenAI and forces prices down could prompt a “much different conversation about these AI labs and the IPOs very quickly”, Brookhart added.

But bullish investors argue that surging demand for AI will allow both labs to thrive even as prices fall, with rising usage more than offsetting lower prices.

Yet the fundamental problem of measuring returns on AI usage remains unresolved. Current systems are akin to judging employees by time spent in the office rather than what they accomplish, said Quanyan Zhu, a professor at NYU Tandon School of Engineering and author of a paper on token economics.

Zhu and other researchers expect some pricing to shift towards results — or service-based — models, where customers pay for completed tasks or agreed outcomes rather than the tokens consumed delivering them.

OpenAI is already testing such an approach, following companies including Salesforce and Sierra. As businesses move towards a more disciplined use of AI, they will likely need to go beyond tokens and work out what AI output is worth to them, said OpenAI’s Chatterji.

“This idea that all tokens are the same is just an artefact of us being able to count them,” he said.

Directors’ Deals: Entain’s new finance chief builds stake

Brazil’s online gambling ban likely to hit group profits

Entain shares are down by half over the past 12 months, which is unsurprising given the wave of regulatory obstacles that have hit the gambling sector in recent months. 

It has made for a bumpy start for new chief financial officer Michael Snape, who succeeded his longstanding predecessor Rob Wood this spring, just weeks before UK remote gaming duty almost doubled to 40 per cent in April. 

Wood’s parting promise was to offset half of the financial impact from higher UK levies partly by using AI to reduce marketing costs. This cost-focused approach has no doubt informed the new CFO’s strategy to protect earnings before interest, tax, depreciation and amortisation of between £910mn and £960mn for the 2026 financial year. 

But the goal took a fresh hit this week after Brazil provisionally banned all online gambling with immediate effect, in a sudden move that rocked investor confidence in the sector. 

While Entain maintained its earnings guidance, the Ladbrokes owner conceded that profit may now sit closer to the “lower end” of its forecast range. Brazil makes up 5 per cent of Entain’s online gaming revenue. 

For Snape, who spent £249,992 on 51,815 shares 10 days before the announcement, the 482p a share he paid now looks a little rich. 

Still, it’s a purchase that was likely made with Entain’s minimum shareholding requirements in mind rather than a case of opportunistic buying. As a newly minted executive, Snape is required to build a stake equivalent to 350 per cent of his £595,000 salary. 

Starting from nil at the time of the group’s 2025 annual report, the finance chief has been quietly adding to his holdings and now owns 64,634 shares. Entain allows its executives five years to build their stakes to the required level.