Cheaper AI Could Make Computing Power More Expensive — Grayscale

cryptonews.ruPublished on 2026-10-12Last updated on 2026-10-12

Abstract

Cheaper AI Could Make Compute More Expensive, Says Grayscale Grayscale's research head, Zach Pandl, explains that cheaper AI models and the rise of autonomous AI agents could paradoxically increase demand for—and the cost of—computing power and physical infrastructure. While competition from cheaper open-weight models is lowering token costs, the expansion of AI usage, especially by agents, may outweigh efficiency gains. AI agents, which perform multi-step tasks autonomously, consume 5 to 50 times more tokens than simple chatbots. Data from OpenAI and others shows agent token usage has grown roughly 14-fold since early 2026, with the most active companies generating over 8 times more tokens than typical firms. Goldman Sachs forecasts a potential 24x increase in token consumption by 2030, largely driven by agents. Even as models become more efficient, overall compute demand will rise if AI usage growth outpaces these efficiency gains. Physical infrastructure—power, data centers, GPUs—remains a bottleneck, as seen in higher-priced compute contracts from providers like CoreWeave. Pandl concludes that cheaper AI and autonomous agents will broaden technology applications, potentially intensifying competition for limited compute resources and benefiting companies that own the underlying infrastructure.

  • Grayscale explained why cheaper AI could increase the demand for computing power.
  • An analyst emphasized that AI agents consume up to 50 times more tokens than chatbots.
  • Goldman Sachs believes token consumption could grow 24-fold by 2030 due to the development of AI agents.

Zach Pandl, Head of Research at Grayscale, explained why the declining cost of artificial intelligence and the development of autonomous AI agents could increase the demand for computing power and the physical infrastructure to support them.

Artificial Intelligence usage is increasing, and token consumption is expected to increase 24x by 2030. Agents, @OpenAI’s "dots," and new lower-cost models could further amplify demand for compute. But the physical infrastructure powering compute remains a bottleneck. Read more on The Stack: https://t.co/A4Pzr7Iv1W

— Grayscale (@Grayscale) October 10, 2026

According to his assessment, the expansion of AI usage could outweigh the growth in model efficiency and create additional opportunities for owners of scarce computing resources.

AI Agents Increase Token Consumption

According to Pandl, competition from cheaper open-weight models is reducing the cost of tokens—units of text that AI processes and generates. For example, the latest Sonnet model from Anthropic reportedly costs up to 30% less for performing a single task.

At the same time, the development of autonomous agents capable of working on user goals around the clock could significantly increase the volume of computations. Unlike chatbots, which answer individual questions, agents perform sequences of actions: analyzing data, processing documents, and preparing results.

According to Grayscale's estimates, such workflows can consume 5 to 50 times more tokens than typical chatbot interactions, depending on the task.

This trend is confirmed by data from OpenAI, OpenRouter, and Similarweb, published in an analytical report by Andreessen Horowitz's crypto division—a16z. According to the data, AI agents use nearly five times more tokens than humans, and their consumption has grown approximately 14-fold since February 2026.

The most active companies generate more than eight times as many tokens as typical enterprises. The growing use of agents also affects the demand for memory and traditional automation tools.

Demand for Computing May Outpace Efficiency Gains

Pandl cites a Goldman Sachs forecast that token consumption could grow 24-fold by 2030, largely due to the proliferation of agents. Even if models become more efficient, overall demand for computing will increase as long as AI usage grows faster.

At the same time, digital services depend on physical infrastructure: electricity, data centers, and graphics processing units (GPUs), whose delivery takes time. As an example, Pandl cited CoreWeave, which reported signing new contracts at higher prices for computing power.

In Pandl's opinion, the declining cost of AI and the development of autonomous agents will broaden the spectrum of technology applications and may intensify competition for limited computing resources. This creates potential advantages for companies that own the relevant infrastructure.

It is worth recalling that, according to Meta CEO Mark Zuckerberg's forecast, billions of people will be using personal AI agents within the next five years.

Related Questions

QAccording to Grayscale, why could cheaper AI models lead to more expensive computing power?

AGrayscale's analyst suggests that cheaper AI models lower the cost per token, making AI more accessible. This, combined with the rise of autonomous AI agents that consume vastly more computational resources (5-50 times more tokens than chatbots), is expected to significantly increase overall demand for compute. This growing demand could outpace the physical infrastructure's ability to supply it (a bottleneck in data centers, GPUs, and energy), potentially driving up the price for computing power.

QHow much more computationally intensive are AI agents compared to chatbots according to the article?

AAccording to Grayscale's analysis, the workflows of autonomous AI agents can consume 5 to 50 times more tokens than a standard interaction with a chatbot, depending on the task.

QWhat is Goldman Sachs' projection for AI token consumption growth by 2030, and what is the main driver?

AGoldman Sachs projects that AI token consumption could grow 24-fold by 2030. The primary driver for this massive increase is the widespread adoption and use of autonomous AI agents.

QWhat physical infrastructure limitations does the article highlight as a bottleneck for AI compute growth?

AThe article identifies the physical infrastructure powering AI compute as a bottleneck. This includes the availability of key components like data centers, graphics processing units (GPUs), and the electrical power required to run them, all of which have long lead times for supply.

QWhich company was cited as an example of rising compute costs due to increased demand?

AThe article cites CoreWeave as an example, noting that the company reported signing new contracts for computing power at higher prices, reflecting the increased demand.

Related Reads

From QR Codes to Physical Goods: How Consumer RWA Verifies Products and Rights

A single QR code can be printed on two bottles and scanned twice to open the same page, so verifying a physical item requires confirming the binding between the bottle in hand and its listed number. Consumer RWA verification has two parts: whether the goods or services are real, and whether the rights recorded in the digital credential are valid. For rights provenance, take a co-branded gift box: authorization documents define the term and scope in which a brand may use the co-branded mark, production and warehousing records show whether the box was stocked by batch, and rights registration states what holders can redeem and who delivers it. Linking all three lets consumers check authorization scope, supply, and their own eligibility. Identifiers, events, and evidence must work together. High-value items can use unique serial numbers, uniform-spec goods can use batch numbers, and courses or events can use session numbers. Events such as warehousing, sampling, outbound shipment, receipt, and redemption should be tied to the corresponding object, with operation time and responsible party retained. Test reports correspond to submitted samples; linking results to specific goods also requires checking the sample, batch, and inbound/outbound records. Ordinary batch goods can combine serialized QR codes with abnormal-scan records, while high-value items may add tamper-evident labels or independent authentication. Trusted records must also reflect current status. If a coffee batch has 500 bags at warehousing, 10 used for sampling, and 100 reserved for other channels, then absent other changes, 390 bags remain available for new rights. The warehousing record shows historical quantity; current redeemable quantity must be updated based on actual occupancy. A query page should at minimum state the goods or service, evidence source, last update time, and whether the credential is currently usable.

marsbit7m ago

From QR Codes to Physical Goods: How Consumer RWA Verifies Products and Rights

marsbit7m ago

10 Years Is Too Slow: Gu Quanquan's Team Challenges Dario, Compressing Cancer Cure to 4 Years

A UCLA professor, Quanquan Gu, has founded Geodesic Intelligence, a 15-person startup that after only five months has released NovaAtom-Lite, an all-atom structure prediction model that outperforms DeepMind's Isomorphic Labs flagship IsoDDE on FoldBench, the toughest biomolecular benchmark. The model predicts the 3D coordinates of every heavy atom in protein, DNA, RNA, and ligand complexes with near-experimental precision. NovaAtom-Lite posted top scores on two core pharma tests: antibody-antigen binding (79.07% success, 31 points above AlphaFold 3 and 3.5 above IsoDDE, with 2.2x AlphaFold 3's accuracy at high-precision thresholds) and protein-ligand binding (top 1 on both RMSD and contact metrics). On the Runs N' Poses benchmark, which separates memorization from genuine physics understanding, NovaAtom-Lite's lead grows on the most novel, out-of-distribution ligands (47.6% vs AlphaFold 3's 24%), showing it learns transferable physical rules rather than memorizing structures. The model also supports test-time scaling—raising sampling from 1 to 100 seeds lifts antibody success from 74.3% to 83.1%—and condition constraints, where a single contact-pair hint boosts success to 91.28% and full pocket hints lift hard-case ligand accuracy by nearly 16 points. It solved two nightmare cases: a wildly flexible 22-residue MERS nanobody loop (0.56 Å error, all 25 samples correct) and the KRAS G13D hidden switch-II pocket (0.98 Å ligand RMSD vs OpenDDE's 17.74 Å). Gu previously led ByteDance's SeedFold, the first model to beat AlphaFold 3 on FoldBench, forcing Isomorphic Labs to respond with IsoDDE. A stronger NovaAtom-Pro is coming. Days after launch, a16z cited NovaAtom-Lite in a report arguing pharma is entering its "GPT-3 moment," naming Geodesic as the sole company highlighted in its industry map. Geodesic's closed loop spans structure prediction, scientific reasoning, and wet-lab validation—aiming to compress drug discovery from a decade to four years.

marsbit22m ago

10 Years Is Too Slow: Gu Quanquan's Team Challenges Dario, Compressing Cancer Cure to 4 Years

marsbit22m ago

Anthropic Misses Out on 'Largest IPO in History'—Who Should Foot the Bill?

Anthropic's IPO has gone from being touted as a potential record-breaking "largest IPO in history" to falling short of that title. Months ago, analysts estimated its IPO could surpass SpaceX, but current data shows it will not. Both its fundraising size and initial market capitalization fall below SpaceX's records of $75 billion raised and $1.78 trillion initial valuation. Anthropic's IPO is priced at $70–78 per share, offering 825 million shares, for a base fundraising of roughly $57.75–64.35 billion. Even with the 15% over-allotment fully exercised, proceeds would reach about $74 billion—still below SpaceX's $75 billion initial raise (and SpaceX's final over-allotment total of $86.25 billion). Anthropic's initial market cap is estimated at $1.26–1.41 trillion, well below both the earlier $2 trillion expectations and SpaceX's $1.78 trillion. However, combined with a $26 billion private placement, Anthropic's total fundraising could approach $100 billion. Three main factors are cited for the setback. First, after Democrats swept both chambers in the U.S. midterm elections, they pushed aggressive AI regulation, with some lawmakers calling for a pause on advanced AI development and a ban on superintelligence. Second, Nvidia founder Jensen Huang resigned as CEO amid a scandal, disrupting the AI supply chain; Nvidia was expected to subscribe about $10 billion of Anthropic's $26 billion private placement, and its instability threatens market confidence. Third, Anthropic's Q3 results show hidden concerns: revenue hit $16.956 billion (up ~11.27x year-over-year), but sequential growth slowed from 144.2% to 46.7%, new quarterly revenue fell from $6.824 billion to $5.401 billion, GAAP operating profit swung from a $974 million gain to a $567 million loss, and adjusted operating margin dropped from 19% to 15%. Notably, reports suggest Huang's ouster involved Democratic figures and media, and that Anthropic founder Dario and the company itself provided funding to related political organizations and donated tens of millions to Democratic midterm campaigns. In this light, Anthropic's dashed "largest IPO" dream is tied to its political bets.

Odaily星球日报1h ago

Anthropic Misses Out on 'Largest IPO in History'—Who Should Foot the Bill?

Odaily星球日报1h ago

Asian chip stocks tumble ahead of Nvidia Q3 earnings

Asian semiconductor stocks fell on Monday ahead of Nvidia's quarterly earnings, as analysts look for signs that the AI spending boom remains on track amid uncertainty over the chipmaker's leadership and future growth. Japanese chip-testing equipment maker Advantest dropped about 10.4%, semiconductor inspection equipment maker Lasertec declined more than 10.7%, and chipmaking equipment supplier Tokyo Electron fell 5.7%. South Korean memory-chip makers Samsung Electronics and SK Hynix lost 5.8% and 4.2%, respectively, while Taiwan Semiconductor Manufacturing Co. slipped 3.8%. The broad decline could reflect caution toward AI-linked stocks ahead of Nvidia's results, though the weakness came amid a wider sell-off in regional equity markets, with Japan's Nikkei 225 down 2% and South Korea's Kospi losing 2.6%. Nvidia is scheduled to release its fiscal third-quarter results on Tuesday, followed by an earnings call. The results will be Nvidia's first since founder Jensen Huang resigned as CEO. Investors and analysts will be watching how Nvidia responds to the AI demand boom while its leadership is in turmoil. Morningstar said it will be looking to see whether the mid-term elections have led to any concern about future demand and whether Nvidia's outlook for the industry has changed. In August, Nvidia forecast fiscal third-quarter revenue of $108 billion, plus or minus 2%, excluding any data center compute revenue from China. Analysts expect Nvidia to report slightly higher quarterly revenue of $109.3 billion, including $101.8 billion from its data center business, according to Visible Alpha consensus figures published by S&P Global Market Intelligence. S&P Global said revenue expectations had risen since the previous quarter on anticipated strength in Nvidia's Blackwell chips. Morningstar analysts also said they expect another "beat-and-raise quarter" out of Nvidia, given strong capex trends among hyperscalers and enterprises, adding they would be stunned if Nvidia provided an earnings miss or weak forecast. Nvidia shares closed Friday down 2.01% at $162.85.

cointelegraph1h ago

Asian chip stocks tumble ahead of Nvidia Q3 earnings

cointelegraph1h ago

Trading

Spot
活动图片