The AI business is coming to the end of a tumultuous year, with another one ahead.
AI models pushed the boundaries of cost in 2026. Vendors had to confront the reality that they still have much to do to contain the AI agents that have repeatedly escaped their sandbox environments, raising alarms around the world about the risks and dangers of the explosively growing technology.
In 2027, the AI market will have to address both problems.
Tokenmaxxing is a 2026 trend that swept up companies starting to use generative AI at scale, with many AI vendors and enterprise customers pushing their employees to max out the number of AI tokens they use. It has been all but abandoned.
Meanwhile, many users turned to cheaper open-weight models from China, helping fuel a price war among the trio of top U.S. model makers: Anthropic, Google and OpenAI. OpenAI cut the price of its GPT-6 Series in half compared to GPT-5.6. Google also slashed the price of its Gemini 3.7 Flash model to half of the previous version's price.
All that underpricing effort showed that model prices are already dropping for enterprises. But token cost alone is only half the story, said William McKeon-White, an analyst at Forrester. Enterprises are not only trying to manage the costs of models but also those of SaaS platforms they use, he said.
“It’s very challenging for organizations right now. It’s going to be less challenging in the next year in part because we’re going to have better tooling, frameworks and systems … and better cost awareness and control,” McKeon-White said.
However, enterprises possessing more AI tools and being more aware of AI costs doesn't necessarily mean that costs will go down; in fact, it's likely that costs will increase because of accelerated use of AI models and agents. Gartner predicts that global spending on AI platforms and models will increase 54.2% in 2027, rising from $64 billion to $99 billion.
Nick Patience, an analyst at Futurum Group, agreed that spending on AI will increase despite model pricing decreasing.
"I see the price per token continuing to fall in 2027, but the overall cost to enterprises will go up because usage will increase,” he said, adding that enterprises will likely focus more on the cost of completed tasks in 2027.
“If you’ve got a relatively cheap model that fails and retries three or four times, that could end up costing more than an expensive one that gets it right,” Patience said. “So, there’s more focus on the quality of the model’s execution of a task than just how many tokens it uses.”
Agents managed better
In 2027, the AI market will be forced to address how enterprises manage agents. Deepak Seth, an analyst at Gartner, said managing agents is becoming a critical area of focus as they flourish throughout IT.
“As agents proliferate, people will have to start thinking about it,” he said. Enterprises will also shift how they use AI agents from saving time in low-risk applications toward using them to deliver quantifiable value or drive business transformation, such as enabling users to build their own applications or eliminating entire departments.
Patience, said: “You will see labs building for things that are agent-first, so things like tool use, computer use, reliability of multistep tasks – those things will matter more than the quality of the chat.”
Better model orchestration and delegation
As agents proliferate, enterprises will focus on orchestrating multiple models rather than relying on one.
“It won’t be a beauty contest among raw models anymore,” Seth said. 'It will be an orchestration battle. The real battle is shifting to the platform control layer, which decides which model handles which task.”
Enterprises might use a premium model such as Anthropic’s Opus 5.5 for complex tasks while routing simpler work to a cheaper Anthropic model like Haiku, raising the question of whether to manage that routing logic internally or delegate it to a provider.
Next year is also when enterprises will make decisions about distributed compute, with some models running in the cloud and others locally to cut costs, amid growing interest in bringing models that sit closer to enterprise data.
“There's an increase in interest in on-premises solutions, even from the hyperscalers,” Futurum Group's Patience said. Enterprises will likely also focus on running models on premises to bring models to where the data is.
Open vs. closed AI
Open-weight AI is gaining enterprise adoption, with models from Western developers, such as Cohere and Mistral, competing with offerings from Chinese developers, such as Alibaba's Qwen and DeepSeek. But proprietary providers still capture a much larger share of metered AI consumption.
In 2027, the extent to which enterprises shift workloads to open-weight models could determine how much pricing flexibility proprietary providers retain.
Paris-based generative AI vendor Mistral added to its open-weight lineup on Oct. 29, releasing weights for Mistral Large 4, also known as Le Chonk. The 1-trillion-parameter multimodal model boasts 52 billion active parameters and targets coding and cybersecurity workloads, with Mistral claiming it beats competing open-weight models on coding benchmarks.
The release comes as more enterprises move open-weight models into production. An October Futurum Group report found that 42% of 200 enterprises surveyed ran open-weight models in production in September, up from 31% in July. Among those users, the share of AI tokens from open-weight models is projected to rise from 34% to 41% within a year.
But adoption hasn't translated into comparable revenue share. Open-weight models accounted for just 5% of enterprise spending on usage-based AI services compared to 36% for Microsoft 365 Copilot and 29% for Anthropic's Claude.
Michael Heinrich, CEO of 0G Labs, attributes the gap to workload allocation: "Companies send routine, high-volume work to open-weight models, while the hardest reasoning goes to proprietary models that can cost 10 to 30 times more per token."
Manish Jain, CEO of Strategic Horizon Research, said enterprises still favor proprietary models for performance, convenience and managed services, while open-weight models demand specialized expertise to run reliably, securely and economically.
Open-weight models give buyers more control
In 2027, enterprises could split more of their workloads between open-weight and proprietary models, using open models for high-volume, lower-cost tasks and closed models for complex reasoning.
Heinrich described this as a hybrid approach: "The common pattern we see is one frontier model call to plan and many low-cost calls to do the work."
However, shifting more workloads to open-weight models also comes with infrastructure and operational costs.
"Companies still pay for hardware and engineering resources, even when GPUs sit idle," Heinrich said. He expects U.S.-based closed-model providers to respond with cheaper small models, their own open-weight releases and more bundling into enterprise agreements.
Jain expects open-weight adoption to expand first in areas such as internal knowledge assistance and document analysis, where data control, cost predictability and data sovereignty matter more than marginal performance gains. Proprietary providers could compete by strengthening their reliability, security, tooling and service guarantees, he said.
Still, Jain said, he doesn't expect a wholesale shift to open-weight AI over the next year. Instead, open-weight alternatives will give enterprises greater negotiating power if proprietary providers raise prices or change their terms.
"Open-weight models give CIOs a credible alternative, and that changes what they are willing to pay proprietary providers," he said.
Robotics and physical AI in 2027
Robotics and physical AI are set to keep expanding in 2027 but unevenly across regions. Growth will be concentrated in the Asia-Pacific region, with AI-related manufacturing and warehouse automation leading the market.
Yueqin Shen, an analyst at Omdia, a division of Informa TechTarget, said that based on Omdia's survey of industrial robotics and automation end users, robotics and physical AI will transform warehousing and logistics over the next year.
Plans for deploying autonomous-guided vehicles (AGVs) and autonomous mobile robots (AMRs) are expected to see the most significant growth, with over half of respondents planning new deployments in the next 18 months, Shen said.
Mobile vehicular robots are autonomous systems used to transport goods and automate material handling in warehouses, distribution centers and manufacturing plants. They typically transport materials around a job site and are used to patrol the perimeter of industrial sites such as mining operations.
Omdia also estimated that revenue from machine vision products in the logistics industry will have the highest growth rate from 2025-2030 (14.5%), largely due to the increasing application of AGVs and AMRs.
Cobots keep outpacing the market
Meanwhile, collaborative robots (cobots) will continue to be a high-growth segment and outperform the market for traditional industrial robots, according to Samantha Mou, an analyst at Interact Analysis.
Cobots are designed for safe use alongside human workers, typically in industrial applications at manufacturing and assembly plants.
“We expect the cobot market to grow by around 20% in unit shipments in 2027, broadly in line with the strong growth seen in 2026,” Mou said.
Three factors are driving that growth:
- AI integration. AI capabilities, more prevalent in cobots, improve performance and open new applications.
- SMB adoption. Smaller companies are “particularly attracted to cobots’ ease of use features,” Mou said.
- Moving upmarket. Higher-payload models and more competitive pricing from emerging vendors are pushing cobots into welding and palletizing. “The growth of cobots is also partly driven by taking share from traditional industrial robots,” Mou said.
Humanoids: More promise than reality
Humanoid robots are unlikely to ship in large numbers in 2027.
While often the most visible face of AI-driven robotics in the media, Omdia's Shen noted they're mostly used for handling, with industrial applications for assembly and precision operations “still limited.”
Paul Miller, an analyst at Forrester, said unrealistic expectations about humanoids need to be tempered, so as not to disappoint investors, early adopters and the broader market.
“Humanoid robots are, undeniably, impressive. But too much emphasis is being placed on them,” Miller said. “We need to hope – and ensure – that the disappointment with expensive, fragile and limited humanoids doesn't spill over and damage the prospects for more pragmatic physical AI form factors.”
Training data is another hurdle.
The world models and other conceptual models that help machines understand their surroundings are often trained on generic data scraped from the internet, low-quality data collected from real environments and synthetic data generated by computers. Higher-quality training data that accurately captures all aspects of the real environment where the machines will operate is more expensive and scarcer.
“There's a place for all these different types of training data, but we need to ensure that the resulting models do more than just look realistic – they have to actually be realistic,” Miller added.
What to watch
Expect 2027 to favor vendors with exposure to electronics, semiconductors and warehousing, as well as those with strong cobot and AI offerings. While humanoids might generate headlines, mobile robots and cobots will produce the volume.
Miller also pointed to world models as an area worth watching.
“We're seeing some of the biggest technology companies competing with well-funded startups to build conceptual representations of the real world that – increasingly – will help any suitably equipped robot work out how to interact with the environment around it,” he said,
Esther Shittu is a TechTarget news writer and podcast host and producer covering AI technology. Kinza Yasar is an AI reporter at TechTarget, with a focus on ethics, enterprise adoption, governance and business strategy. Scarlett Evans is a freelance writer with a focus on robotics and emerging technologies.