Dansk resume af analyse fra Authers, FT:
Analysen handler om, hvordan udviklingen i AI-markedet kan måles gennem forbruget af såkaldte tokens – de små datapakker, som store sprogmodeller bruger. Tokenforbrug fungerer nærmest som en slags “AI-valuta”, fordi det viser, hvor meget virksomheder og brugere reelt betaler for at anvende AI.
Den centrale pointe er, at Silicon Data Token Expenditure Index er faldet kraftigt efter tidligere rekordniveauer. Det betyder dog ikke nødvendigvis, at AI generelt er blevet billigere. Faldet kan skyldes flere ting: lavere priser, at brugerne flytter over mod billigere modeller, eller at efterspørgslen efter dyre AI-løsninger er blevet svagere. Grafen i original analysen nedenfor viser, at token-udgifterne steg markant i starten af perioden, men siden er faldet tilbage fra toppen.
Analysen peger på, at høje tokenpriser kan få brugere til at vælge billigere alternativer, blandt andet kinesiske AI-modeller. Alligevel mener den bullish investor Louis Navellier ikke, at investorer bør frygte et større kollaps i AI- og datacenteraktier. Hans argument er, at ordrebøgerne hos AI- og datacenterselskaber er meget stærke og rækker flere år frem, hvilket tyder på, at datacenterboomet kan fortsætte mindst frem mod 2029.
Et andet vigtigt tema er, at de store teknologiselskaber – Microsoft, Amazon og Meta – bruger enorme beløb på AI-infrastruktur. Den anden graf viser, at næsten hele deres frie pengestrøm nu går til investeringer, især chips og datacentre. Det ændrer investorernes syn på techsektoren, fordi de store selskaber i stigende grad ligner kapitaltunge industrivirksomheder.
Samtidig har chipaktierne klaret sig langt bedre end de store hyperscalere. Den tredje graf viser en tydelig divergens, hvor halvlederaktier er steget kraftigt, mens hyperscalerne er sakket bagud. Morgan Stanleys Michael Wilson vurderer, at denne forskel ikke er holdbar, og at chipaktierne kan stå foran en korrektion, mens hyperscalerne stabiliseres.
På længere sigt er analysen stadig positiv over for AI’s produktivitetspotentiale. En undersøgelse fra Boston Consulting Group viser, at virksomheder med højere tokenforbrug også har stærkere omsætningsvækst. Virksomheder i den højeste tokenforbrugsgruppe har markant højere vækst end dem med lavest forbrug. Pointen er, at AI kan gøre vidensarbejdere mere produktive – på samme måde som samlebåndet gjorde produktionen mere effektiv.
Men analysen advarer også mod overdreven optimisme. AI-gevinsterne kommer næppe hurtigt og bredt overalt. De største fordele vil først opstå dér, hvor gevinsterne kan retfærdiggøre de store omkostninger til computerkraft, strøm og infrastruktur. Derfor bliver udviklingen sandsynligvis ujævn: nogle virksomheder og sektorer vil få store fordele, mens andre vil være langsommere til at få værdi ud af teknologien.
Kort konklusion:
AI-boomet er ikke nødvendigvis slut, men den nemme og lodrette optur er sandsynligvis ovre. Markedet begynder nu at skelne mere mellem hype og reel økonomisk værdi. Fremover bliver det afgørende, hvilke selskaber der kan omsætte AI-investeringer til konkret produktivitet, omsætningsvækst og rentable forretningsmodeller.
—————–
Uddrag fra Authers, FT:
More Than a Token |
Amid the ever-growing array of metrics of the artificial-intelligence boom, one relatively useful, straightforward measure tracks how much is being spent on large language models. Why not just count the total number of tokens — the tiny packets of data that are the building blocks of AI models — people buy to use them? It’s as close to an AI currency as we currently have, and while it might not ultimately prove the best gauge of the technology’s trajectory, it’s a real-time gauge of AI companies’ pricing power. On this basis, the Silicon Data Token Expenditure Index’s plunge from record highs offers signs of trouble brewing beneath the surface:
Bloomberg News colleagues Jan-Patrick Barnert and Michael Msika point out here that this doesn’t mean that AI is getting cheaper. As the gauge blends prices and usage, a dip could reflect either falling list prices, a shift in demand toward cheaper models, or a softening in how much buyers are willing to pay. Only weeks ago, the index was trading at twice its level from late last year, even as the price of a single token had plunged by more than 90% since 2023.
For now, veteran bullish technology investor Louis Navellier argues that high token prices could be pushing users toward lower-cost alternatives, particularly Chinese offerings, and dampening enthusiasm for all things AI. What does this signal for the broader AI trade in the near-term? Not a great deal, he argues:
I do not want investors to worry about the selloff in AI and data center-related stocks. In my opinion, the big news is that the order backlogs for AI and data center-related companies surged to nearly a three-year backlog, so the data center boom is expected to persist at least through 2029.
Meanwhile, Navellier’s expectations align with the sense that investors looking to find AI winners will ultimately gravitate toward hyperscalers such as Microsoft Corp. (currently down 28% from the all-time high it set last October), Amazon.com Inc., and Meta Platforms Inc., given the strength of their core businesses. As Bank of America Corp.’s Savita Subramanian demonstrates in this chart, the AI arms race of the last two years has had an extraordinary effect on their free cash flows. Virtually all of the cash these companies produce is now going to capital expenditures (meaning chips):
As investors have grasped that these companies now depend on capital as much as old-line industrials, eliminating a key perceived advantage that the tech sector has enjoyed for decades, so their share have stalled. Attention has naturally turned to the massive sums they have to pay to the chipmakers of the SOX semiconductors index:
Morgan Stanley’s Michael Wilson argues that chips’ momentum will fade as investors grow increasingly wary of their ballooning valuations, rising competition, and potential overcapacity — and remember the potential payoff from massive AI investments for hyperscalers:
I expect the hyperscalers now to stabilize; that’s what’s been going on in the last couple of weeks, and the semiconductor stocks are going to correct… That’s a good development. You can’t have this divergence continue; it’s not sustainable.
The likelihood remains that AI will positively affect productivity, which should ultimately drive the technology’s attractiveness — even if competition pushes down prices for cost-sensitive consumers. A recent Boston Consulting Group paper by Matthew Kropp and his team analyzed token consumption at more than 100 public technology companies with more than $500 million in trailing 12-month revenue. It concluded that tokens multiply what knowledge workers can produce, much as the assembly line did for manufacturing:
The firms pulling ahead are redesigning their processes around AI, with people applying expertise to direct and improve the system. These companies operate at a pace and scale that human-only organizations cannot match.
Ultimately, they find that if the winners get it right, they’ll create workflows that improve organically as AI becomes better, faster, and cheaper, and the rewards could be significant. The following chart shows that the pattern among BCG’s survey is already consistent with productive token use translating into an advantage:
For all AI’s promise, the path to realizing significant productivity improvements isn’t straightforward. Citadel’s Frank Flight argues that the journey will be more selective and cost-conscious than markets once assumed. That is relevant for asset prices:
The key variable is not productivity alone, but the elasticity of demand for the tasks and services whose costs are falling. Where demand is elastic, AI-enabled efficiency gains can expand output enough to raise demand for complementary factors of production, such as labor.
Over the medium term, Flight notes that improvements in model efficiency and power infrastructure can offset some of today’s physical constraints. He argues that markets, however, should avoid anchoring themselves to a world in which AI is ubiquitous, frictionless, and immediate — even though that is close to what eventually happened after the introduction of the internet.
Ultimately, a more plausible path is one of uneven diffusion: Frontier deployment concentrated where the returns justify the computing power, broader adoption shaped by cost and capacity, and asset prices periodically forced to reconcile ambition with physical constraints. Put differently, the period of vertical ascent is probably over.
2 md. adgang for
2 x 49 kr.
Få straks adgang til denne artikel og derefter 2 måneder til alle artikler på ugebrev.dk
- Alle artikler på ugebrev.dk
- Om investering, finans, ledelse, samfundsansvar, life science og Bestyrelsesguiden.dk
- Daglige nyhedsmails med nyheder og analyser
Tilbuddet gælder til 31. juni 2026. Abonnement fortsætter til normalpris på 249 kr. efter bindingsperiode på to måneder. Opsig når du vil - til udgang af den anden måned. Tilbud gælder kun, hvis du ikke har haft abonnement på ØU udgivelser de seneste tre måneder
Allerede abonnent? Log ind her
