Gartner forecasts open-weight models will run 90% of enterprise agent steps by 2030
Editorial Team
••📖7 min read
Gartner forecasts that open-weight models will execute 90% of enterprise operational agent steps by 2030. Its cost estimates favor routine, high-volume work, but infrastructure costs and the choice of comparison model leave enterprise savings to be proved.
Gartner forecasts that open-weight models will execute 90% of enterprise operational agent workflow steps by 2030.
Its best-case twentyfold cost advantage over frontier models narrows to roughly five- or sixfold after longer outputs and retries.
Local deployment offers more control over workflow data, but serving infrastructure and staffing can change the economics.
Gartner, the technology research and advisory firm, expects open-weight models—models whose trained parameters can be downloaded—to execute 90% of enterprise operational agent workflow steps by 2030. The reported forecast concerns routine executions such as classification, tool calls and data validation, rather than entire workflows completed independently.
That would change which models enterprises pay to use repeatedly. AI agents combine models with software tools to carry out tasks; they can use an inexpensive specialist for routine operations while reserving the most capable general-purpose, or frontier, models for harder reasoning. A model's share of executions therefore does not establish its share of spending, revenue or business value.
Gartner’s company-reported forecast: open-weight models will execute 90% of enterprise operational agent workflow steps by 2030. This is not an observed adoption rate. Source: The Register.
Three cost comparisons, not a twentyfold budget cut
The projection comes from Gartner's paper “Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps,” published last week, according to the report quoting it. Its authors, Gartner analysts Mark Petty and Raymond Paquet, argue that nine in ten agent workflow steps are routine and that open-weight models can handle them at comparable quality for less money.
Their reported comparisons use different baselines:
SK hynix bought existing Penguin Solutions preferred shares from an SK Telecom investment vehicle. The deal brings director nomination rights and a convertible stake in an existing AI infrastructure partner, while sending the purchase price to the seller.
Frontier comparison accounting for longer outputs and retries
Roughly five- to sixfold saving per task
Those are analyst estimates. The published accounts do not name the models, benchmark, serving configuration or full costing method behind them. In particular, using a frontier model for every step is a different baseline from using the cheapest suitable closed model. The twentyfold figure does not establish a twentyfold reduction in an enterprise's total agent operating bill.
Gartner expects the strongest adoption in domain-specific models handling a small number of highly stable steps millions of times a month, a second account of the paper says. Its proposition is most concrete for repetitive, high-volume work, rather than for replacing every general-purpose assistant.
Control depends on where the model runs
Petty and Paquet place greater strategic weight on operational knowledge than on price. They argue that years of workflow history help explain what an acquirer pays for a software-as-a-service company, and that sending routine work to a frontier-model application programming interface, or API, moves some of that information beyond the company's perimeter.
Downloadable weights permit local deployment and adaptation. They do not make all components open source or all uses unrestricted. A policy paper on AI openness distinguishes access to weights from access to code and training data, and notes that licenses can restrict use and distribution.
That paper also documents cloud providers hosting open-weight models. The practical control question is consequently where execution occurs and who handles the data, not simply whether weights are downloadable. Local deployment can reduce dependence on an external API; choosing an open model does not by itself establish that workflow information stays inside the enterprise.
The surrounding software matters too. An agent “harness” manages execution, supplies context, routes tasks among models and applies guardrails. The accompanying analysis argues that this layer increasingly determines product value as models become more interchangeable. It also points to European AI lab Mistral's open-weight Mistral Large 4, also called Le Chonk, as an option for organizations prioritizing sovereignty.
Gartner still assigns frontier models a role in research and development, with open-weight models putting capabilities into routine operation. Its capital-expense versus operating-expense analogy describes that division of labor; it is not evidence that frontier-model demand disappears.
Earlier research supports specialists, with substantial caveats
A June 2025 position paper by researchers affiliated with NVIDIA Research and the Georgia Institute of Technology made a similar case for small models handling repetitive agent tasks, with larger models retained where general capabilities are essential. It drew partly on NVIDIA's own model and inference-software work, alongside examples including Microsoft's Phi models and Salesforce's xLAM-2-8B tool-calling model.
Small models and open-weight models are different categories: one describes deployment size, the other access to trained parameters. Neither category alone guarantees suitability for a particular agent task.
The paper cited a 10–30-fold inference-efficiency advantage for a 7-billion-parameter model over models with 70–175 billion parameters, expressed in latency, energy consumption and computational operations. That is not an enterprise dollar-cost benchmark. More directly relevant to deployment, its appendix estimated different replacement potential across three agent systems:
System examined
Function
Estimated share of large-model queries replaceable by specialized small models
MetaGPT
A framework emulating roles in a software company
About 60%
Open Operator
Workflow automation using tools and services
About 40%
Cradle
Computer control through screenshots and simulated user actions
About 70%
These were the authors' assessments of potential substitution, not measured production replacement rates. They identified harder architectural reasoning, sustained context, changing interfaces and unstructured errors as limitations in the respective systems. The estimates concern small-model substitutions in selected agents, not Gartner's population of enterprise operational steps, and cannot confirm or refute its 90% forecast directly.
The authors also acknowledged that centralized model serving could retain an economic advantage. Specialized endpoints can be harder to keep fully utilized, while infrastructure setup and hiring and retaining staff add costs often omitted from inference comparisons. They called the economics case-specific. The policy paper likewise warns that the benefits of open weights depend on access to computing resources, data and skilled talent.
Company-reported research estimates from Belcak et al., affiliated with NVIDIA Research and Georgia Institute of Technology. These are potential small-model substitutions, not measured production rates or a test of Gartner’s forecast. Source: Small Language Models are the Future of Agentic AI.
The next decision is whether the workflow needs an agent
Gartner's own earlier guidance makes model selection only part of the decision. In June 2025, it forecast that more than 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. That is a forecast, not an observed cancellation rate.
The firm also warned that many purported agent use cases did not require agents. It recommended agents where decisions are needed, automation for routine workflows and assistants for simple retrieval, while noting that integration with legacy systems can require costly changes.
For buyers, the next test is therefore twofold: whether an agent adds value over simpler automation, and which model mix delivers that value at the lowest total cost. A useful comparison must hold task quality constant and count retries, failures, human intervention, integration and serving costs alongside model charges. It must also establish where workflow data travels. Gartner's forecast identifies a possible destination; those deployment results will determine whether the cost and control argument holds.
Google plans to switch off Gemini and Search for school-managed K–12 student accounts unless administrators enable them. Schools will gain control over access, but Google says keeping Search without its AI features will not be an option.
Anthropic is expected to exceed 3.5 GW of computing capacity in 2026 and aims to double it in 2027. Its filing shows $895.4 billion in commitments, substantial financial resources and an immediate cost from scarce capacity.
Independent tests reproduce MIT researchers’ repair of Qwen’s repetitive reasoning. The edited model still beats a stronger unedited configuration, but by much less; broader claims about creating capabilities remain unproven.
FTAI Aviation is financing a planned 100-turbine production run in 2027, while ProEnergy has a 650 MW equipment order from Crusoe. Converted aircraft engines offer another supply route for data-center power, but delivery targets, affiliated financing and unproven operating economics qualify the opportunity.
Meta is marketing dollar bonds in a plan to borrow at least $20 billion, including an expected euro offering. Capital spending exceeded operating cash flow last quarter, but the debt is part of a wider financing strategy that also includes equity offerings and a proposed non-voting share-sale program.
A nonbinding opinion targets real-time retrieval and recognisable reuse of journalism, while rejecting a monopoly over news facts and recommending that questions about model training be dismissed as hypothetical.
A study of Lovable projects found passport details, customer records and private messages accessible without login, raising questions about database permissions and the limits of automated security checks. Its sample does not establish a platform-wide leak rate.
CrowdStrike is extending Falcon Guardian into Claude Cowork’s virtual machine and AWS’s AI guardrail layer. The Cowork beta addresses a visibility gap, while existing AWS safeguards and unfinished acquisitions limit what buyers can infer about the broader platform’s capabilities.
Oman and UK-founded Yamna have agreed to cooperate on a Salalah AI data centre whose first phase is expected to exceed £1 billion. The developer brings a hydrogen-project foothold and an earlier Brazilian data-centre proposal, but financing, customers and delivery dates remain undisclosed in the announcements.
War Thunder will add Nvidia’s DLSS 5 neural rendering in mid-November. The upgrade changes lighting and materials, with access and performance depending on hardware, streaming and other rendering features.
Samsung’s planned second Taylor fab reportedly has its output booked for AI-memory base dies, mostly for AMD, with production targeted for 2030. The orders could underpin investment, but Nvidia’s reported TSMC sourcing requirement limits the case for a broader foundry breakthrough.
Nordic Semiconductor is sampling nRF71 chips combining Wi-Fi, Bluetooth LE and Thread, with an AI accelerator in the higher-end variant. Its claimed standby-power improvement targets battery devices, but software maturity and real-world energy use remain key tests.
The operator of ao3.life says authors submitted more than 2,000 takedown requests after the site repackaged fanfiction and sold AI audiobooks during AO3’s outage, exposing the gap between restoring access and obtaining permission.
CrowdStrike has added an AI Kill Switch to Falcon Guardian that blocks agents’ model access through connected gateways. Broader containment across endpoints, identities and applications remains planned, making deployment coverage the key test of the new control.
DigitalOcean raised its 2026 revenue forecast to $1.208–$1.211 billion after third-quarter sales grew 40%. Larger customers are driving the expansion, while falling operating margins and rising equipment obligations complicate the earnings beat.
A leaker claims Intel may retire the B580, launched at $249, to favor professional AI products. Shared chips make reallocation possible, but no discontinuation notice has been reported, and the hardware can serve both markets.
The independent games marketplace is widening mandatory AI declarations and search filtering after reporting a surge in new game pages. But filtered results and deleted terabytes do not establish how much of its catalog is AI-generated.
A false authority appeared in a senior prosecutor’s reasons for continuing a child’s prosecution. The High Court rejected the challenge, while criticising the explanation for the citation and other aspects of the CPS’s handling of the case.
DeepSeek’s reported eight-trillion-parameter model plan accompanies a push toward Huawei hardware. Joint software tools and a larger reported fundraising show progress, but chip supply, training performance and deployment costs remain unresolved.
