Logo
Become a Client

A century ago, electricity moved into factories and homes far faster than anyone could reliably deploy or meter it. The wiring went in first, then came circuit breakers, safety codes, and accurate billing. Mostly after fires, blackouts, and disputed invoices forced the issue. Enterprises are living a corporate version of that same lag with agentic AI right now, and IDC’s new Perspective, Crossing the Agent Economy’s Cost Governance Gap, puts hard numbers on exactly how wide that gap has become.

How much is already wired into the business

The adoption number alone should settle any lingering debate about pilots versus production. According to IDC’s Future Enterprise Resiliency and Spending Survey (FERS Survey) Wave 4 conducted in July 2026, 95% of enterprises worldwide now run at least one company-funded agent-enabled workflow in production, and the average organization is running roughly 11 of them in different functional areas. This is not an early-adopter statistic. It is an installed base to be managed and extended, and it’s concentrated most heavily in IT ops and software development (71% of organizations), followed by customer service/support (43%), and supply chain and procurement (36%).

The spend backing this growing footprint is real money in action today, not a rounding error to be modeled later. Enterprises with visibility into their own agent costs (most do) report average monthly spend on agent inference and related orchestration services of $117,558, well over $1 million a year annualized, and IDC’s forecasts for agent adoption show this is the early chapter of a much longer story. The population of active agents used by organizations is projected to reach 2.5 billion by 2030, nearly 80 times what it was in 2025. Those agents will be completing 459 trillion actions a year versus 48 billion today. None of this is justification for slowing down, however. The innovation case for agentic AI is real, and enterprises have proven they can carry the initial load. The question IDC is now addressing is whether the meters and breakers enterprises rely on are keeping pace with the expansion and what tech leaders can do to avoid the bad outcomes.

The overrun is a visibility problem, not a discipline problem

They are not keeping pace yet, and the data shows exactly where the gap sits. 67% of enterprises ran over their agent spend budget by more than 10% in the past 12 months, including 24% that ran significantly or extremely over forecast. Breaking that overrun down by severity tells the real story: 43.1% ran moderately over budget by 10% to 25%, 18.5% ran significantly over by 26% to 50%, and 5.4% blew past forecasts by more than 50%. The functional area leading agent adoption, IT and software development, is also the function most often cited for the worst overruns, which suggests enthusiasm and budget discipline are moving in opposite directions inside the same teams.

The consequences are not abstract. Roughly half of over-budget organizations deferred other important IT projects to stay within budget, and 43.1% proactively reduced headcount to compensate. Agent overruns are already crowding out other technology investment and, in a meaningful share of cases, jobs. Underneath it all sits a genuine paradox: 61.8% of organizations rate their own cost governance as “defined” or “optimizing,” yet only 45.4% have real-time dashboards tracking token consumption and cost by workflow. An area where two thirds of organizations believe their governance is mature and two thirds also missed budget is an area where self-assessment and outcomes have stopped tracking each other.

IDC created a framework to attack this exact issue. In Effective Agent Cost Management: A Practitioner’s Framework for Planning, Sourcing, Implementation, and Operations, we offer a short list of fixes that will matter more than any anticipated model price cuts:

  1. Name a single accountable owner for agent spend, per workflow, before the next budget cycle, not after the next overrun.
  2. Replace the monthly invoice review with real-time cost visibility. A once-a-month snapshot cannot manage a cost surface that moves by the hour.
  3. Fund cost governance as an engineering discipline, with engineering time attached, since the controls that work best (model routing, prompt caching, agentic loop guardrails) are built, not written into a policy memo.
  4. Consider capping any single vendor’s share of inference spend, generally below 40%, and keep a second option warm as pricing and terms keep shifting quarter to quarter.

Cost is one thread, not the whole fabric

It would be a mistake to treat agent cost as a standalone finance problem, and IDC’s companion framework, Autonomy on a Leash: A Framework for Agentic AI Governance , explains why. As my colleague Duncan Brown puts it, “organizations that wait to implement governance until after deployment are not avoiding cost. They are only deferring it and compounding their risk while they wait.”

IDC’s framework maps 12 IT domains, including FinOps and cost management, against five operational governance pillars: explainability and auditability, identity and access control, human oversight architecture, systemic agent governance, and continuous monitoring. Cost sits primarily under continuous monitoring, but it never sits there alone. The same organizations running over budget are also the ones where non-human identities already outnumber human ones, and where only 44% of organizations have implemented a policy for agent identity despite 92% calling it critical.

This is the deeper point underneath the spend numbers. An overrun and a governance gap are usually the same failure, seen from two different desks. IDC’s research on organizational governance maturity finds that enterprises with real accountability models, documented oversight, and real-time observability are on track for operating margins up to 15% higher by 2029 than those chasing AI productivity gains alone. Governance, built well, is not the brake on the innovation described above. It is the breaker box that lets an enterprise run agentic AI at full power without finding out where the short circuit was only after the fire.

Navigate the next move with confidence

None of this is an argument for slowing down incorporation of agentic AI into the business. It’s an argument for tech leaders to take the central role in scaling it well, before finance, legal, or a public incident does it for them. IDC created its two frameworks for this exact leadership moment.

Effective Agent Cost Management: A Practitioner’s Framework for Planning, Sourcing, Implementation, and Operations gives you the concrete controls to build a cost plan your CFO can sign off on.

Autonomy on a Leash: A Framework for Agentic AI Governance gives you the governance architecture to build a program your board can stand behind. Pull both apart, map them against your own agent population, and assign owners and dates to what they recommend.

The tech leaders who do that now are the ones who end up wiring their business for agentic AI safely, rather than explaining after an incident why they didn’t. IDC built these frameworks, and the research behind them, to be a trusted advisor for exactly this kind of work, from the first pilot through enterprise-wide rollout. Bring us the plan you’re building, and let’s navigate your next move with confidence.

Rick Villars

Rick Villars - Group Vice President Worldwide Research

Analyst Profile
Rick is IDC's leading analyst guiding research on the future of the IT Industry. He coordinates all IDC research related to the impact of Cloud and the shift to digital business models across infrastructure, platforms, software, and services. He helps…

Between 2023 and August 2025, European IT services buyers fundamentally reordered their priorities. AI capabilities jumped from sixth place to first. Digital expertise became foundational rather than leading. The shift wasn’t gradual. It signaled a market that had moved into new territory.

IDC’s survey of 700 European organisations in August 2025 captured this reordering of service provider selection criteria. When compared with the same evaluation criteria from 2023, the changes were clear. European firms were demanding AI capabilities as a core competency.

AI capabilities ranked sixth in 2023. By August 2025, they ranked first.

Three years ago, when European IT services buyers were asked what mattered most in selecting a strategic services partner, AI and generative AI capabilities ranked sixth on their priority list. By August 2025, they had moved to the top. This reflected a real shift in how European organisations had come to think about the value their partners could deliver.

Most organisations had experimented with AI tools by 2025. Many had run pilots. Few had successfully redesigned core processes to use AI at scale, which is where they needed their services partners to step in and help them navigate that gap. The implication for services providers was that AI competency could not be a future roadmap item. Buyers expected to see real AI capabilities embedded across the services portfolio, and for EMEA services providers, this had become a baseline expectation, not differentiation.

The third leading criterion in the August 2025 survey revealed something practical about buyer psychology: European organisations needed their IT services partners to act as trusted advisors on what was coming next. Firms often have limited internal bandwidth to stay informed about emerging technologies, to evaluate competing options, and to understand which trends genuinely mattered for their industry versus which were hype. They needed technology partners who could close that gap.

The buyer conversation had fundamentally shifted. They weren’t just asking “Can you build this?” They were asking “What should we build? What’s the market doing? What’s coming next, and what does it mean for our business?” Services vendors who could articulate a credible point of view on technology strategy, backed by real market insight, and more importantly, experience, had a competitive advantage.

Industry expertise and innovation rounded out the top selection criteria

The August 2025 survey showed industry expertise and a proven track record of helping clients innovate completed the top five selection criteria for European services buyers. These weren’t new requirements, but by 2025, their context had shifted. What counted as innovative five years earlier often didn’t anymore. Industry expertise that was purely historical wasn’t sufficient either.

European buyers were looking for services partners who combined deep industry knowledge with current thinking about how that domain was being reshaped by AI, regulatory change, and competitive disruption. For services vendors, the implication was clear: industry specialists needed to understand how AI was changing their sector, which business processes were good candidates for AI-led transformation, and why generic industry knowledge wasn’t a sufficient competitive base anymore.

What’s changed since then

The shift between 2023 and August 2025 was clear, but the services market doesn’t stand still. The real question is what happened in the 12 months that followed. Did European IT services buyers continue to demand more from their providers? Did AI capabilities stay at the top of the evaluation criteria, or have buyer priorities shifted further? Have services vendors made the repositioning moves that seemed necessary in late 2025? Have new evaluation criteria emerged? And critically: how have these changing priorities affected actual vendor selection decisions across EMEA?

These aren’t abstract questions. For services vendors, the answers determine whether your current positioning is ahead of the market or reacting to it. For services buyers, they determine how you should be evaluating partners right now.

On 6 October, IDC is presenting updated research that answers these questions with fresh data on how EMEA services buyer requirements have evolved. If you’re involved in selecting IT services partners, or if you’re positioning your services business to win in both today’s and tomorrow’s market, this is when you’ll get the insights you need.

Join the webinar: What will EMEA services buyers demand from their partners in the AI era?

This blog draws on IDC’s EMEA IT Services Survey conducted in August 2025 (n=700, Europe).

Matthew Wilkins

Matthew Wilkins - Research Director, European Services

Analyst Profile
Matthew Wilkins is a research director in IDC's Global Services Insights research team. Based in London, Matthew leads IDC's European Services research program, focusing on the trends and market dynamics impacting professional service providers in Europe, including the disruption and…

A recent Forbes roundtable put eight robotics and supply chain experts in front of the exact question IDC’s modeling on the FCC’s new robotics rule was built to answer: does this restriction actually protect American robotics, or does it just look like it does? Forbes Senior Contributor John Koetsier’s panel landed on cautious, qualified support. Our modeling says the panel was directionally correct but lacked a significant detail IDC reporting uncovered: the category the rule was designed to protect, humanoid robotics, takes the deepest hit of any tracked segment through 2030, while the category assumed most exposed rides it out best. We laid out that full model, category by category, in a companion piece last week. This article checks three of the panel’s individual arguments against IDC’s own research, and asks the question that’s left over: if “protected” doesn’t mean “safe,” what does?

The Gap Between the Policy’s Intent and Its Effect

The FCC’s July 2026 rule blocks new equipment certifications for foreign-made mobile robots, including humanoids and quadrupeds, on national security grounds. The instinct is to read that as a straightforward win for US builders and the humanoid category generally. The reason it isn’t is what the rule can’t reach: the motors, actuators, batteries, and rare earth materials inside every unit, most of which still trace back to China regardless of where final assembly happens.

That’s the gap operations leaders now have to plan around. A trade rule can restrict where a robot is assembled. It can’t, on its own, restrict where the components inside it come from. Treating a “Made in America” label as a supply chain risk mitigant is the mistake this restriction quietly exposes.

What Did IDC Say?

Ryan Reith, group vice president at IDC, and Navkendar Singh, associate vice president at IDC, were two voices in Forbes’ roundtable, credited alongside the other contributors quoted in the piece. Their contribution walked through the scale of the rule’s financial impact, which robot category bears the brunt, why the country of assembly doesn’t settle the question, who stands to benefit as volume shifts elsewhere, and how vendors are likely to respond as the rule takes hold.

So what did the other panelists say, and how does IDC research support their perspectives?

How IDC Research Supports Panelists’ Perspectives

A leader at a robotics component supplier argued the rule serves a dual purpose: a national security measure today, and an incentive for domestic production tomorrow. IDC’s own research on Americas manufacturing (IDC #US53528026) confirms reshoring is real and accelerating. Still, the same research is candid about the gap between incentive and outcome: workforce shortages, hidden transition costs, and business cases that are “often more fragile than anticipated” are the actual execution risk, not the policy intent.

A leader at a robotics hardware and optics firm argued that trust, security, and long-term vendor support must come before country of assembly for enterprise buyers evaluating humanoid robots for critical operations. IDC’s own vendor assessment of autonomous mobile robots (IDC #US53016726) points in the same direction, independently advising buyers to weight vendor stability and multi-year service roadmaps over hardware price.

A third panelist called the rule “a sensible first step,” designed carefully enough that US companies could keep sourcing the best global sensors, motors, and computing while domestic supply grows around them. IDC’s own supply chain research (IDC #US50873823) points to a different outcome: what gets called “made in America” is usually “assembled in America,” with parts still coming from overseas, and IDC expects nearshoring activity to concentrate on final assembly rather than components for the foreseeable future.

What This Means for a Physical AI Business Case

The timing sharpens the stakes. IDC’s most recent research puts physical AI as the #2 AI investment priority for the next two years, just behind generative AI assistants (IDC #AP54804326, August 2026). “For operations leaders, this means governing it well from the start, extending vendor accountability across models, infrastructure, safety states, and human override,” says Stephanie Krishnan, associate vice president, Manufacturing and Supply Chain at IDC.

The Actual Takeaway

IDC’s own framing in Forbes says it more sharply: the market isn’t simply helped or hurt by this rule; it’s splitting into two tracks: mass-market categories that keep growing regardless, and higher-value categories where “the appearance of protection outpaces the reality of supply chain independence.” For a physical AI investment case, the practical version of that split comes down to one unresolved variable: whether a vendor’s component supply and service commitments hold up if the rule tightens, not what’s stamped on the finished unit.

Read the full expert roundtable on the FCC’s rule at Forbes: 8 Experts Weigh In On The FCC Foreign Robot Ban: Good Or Bad?

For IDC’s full category-by-category modeling behind this piece, including the cost forecast and regional breakdowns, see What Does the US Robotics Ban on Foreign Imports Really Mean?

Ryan Smith - Content Marketing Director - IDC

Ryan Smith is the Director of Content Marketing at IDC, where he leads brand-level content and social media strategy, aligning research insights with compelling storytelling to engage technology decision-makers. With a background in both IT and marketing, Ryan brings a unique blend of technical understanding and creative strategy to his work. He’s also a seasoned storyteller, speaker, and podcast host who believes the right message, told the right way, can drive both trust and transformation.

The smartphone market has changed character this year. Units are dropping while prices are climbing sharply, with consumers expected to absorb the cost.

According to IDC’s Worldwide Quarterly Mobile Phone Tracker, worldwide smartphone shipments will fall 16.7% in 2026 to just over 1 billion units.

That is a sharp downgrade from the 13.9% decline we forecast only one quarter ago, and it is the steepest annual contraction the industry has ever recorded. What makes this moment unusual is that the market is shrinking and getting more valuable at the same time. Total market value will still grow 6.3% to $613 billion, because higher prices are now doing the heavy lifting that volume once did.

Why did the outlook get worse?

The memory shortage, which started in late 2025, is striking hard in the second half of 2026, with shipments forecast to drop by 27.2% YoY. NAND and DRAM costs continue to rise, up over 300% YoY, and vendors are running out of options to absorb the increased costs. As memory prices are expected to continue increasing until at least 2028, vendors are adapting their portfolios to a permanently higher cost structure. The 173 million smartphones below $100, which shipped last year, are facing an existential crisis. Android players focused on low-end devices, which were already operating on razor-thin margins, are cutting low-end models and pushing a higher-end product mix. In Q2 2026, this segment saw an almost 60% YoY drop and is expected to fall faster in the second half.

How much are smartphone prices expected to increase?

This is where the forecast has moved most. The average selling price of a smartphone will reach $581 in 2026, up 27.6% in a single year and revised upward from the $550 projected last quarter. More brands are passing the increased cost to the end consumer, with prices rising faster than expected. The premium end will remain more resilient to the price hikes as long-term interest-free financing options are more prevalent in developed markets like the US and UK. The mass market does not hold up, especially in emerging markets, which are expected to drop over 20% this year. This is why the unit decline continues to deepen while the value line keeps rising.

“

The memory tsunami that we warned about is now hitting the market in full, and consumers are starting to pay the AI bill. The components that make AI possible are the same ones in short supply, and their cost is being passed straight through to the shelf. Average selling prices are up 27.6% this year and will keep rising well into 2027. The era of the cheap smartphone has ended. From here, the winners will be the vendors with the scale and supply leverage to hold demand at prices consumers have never had to pay before.

Who is winning the crisis, iOS or Android?

The crisis is not hitting everyone equally. Android bears almost the entire decline, falling 24.3% in 2026 as its most exposed vendors retreat from the entry tiers they can no longer serve profitably. Android share drops seven percentage points in a single year. iOS share moves in the opposite direction, increasing almost four percentage points from last year to a record-high 23.6% share, as shipments remain relatively resilient, down just 1.3% YoY in 2026. HarmonyOS sees strong growth off a small base, nearly tripling to 51 million units in 2026 as Huawei maintains a disciplined pricing strategy, taking full advantage of the crisis to gain share in China as the rest of the Android market contracts.

Is there any segment that is still growing?

Almost nothing in this forecast grows, except foldables. The category will grow 12.6% in 2026 to 22.9 million units, then accelerate to 18% growth in 2027, reaching roughly 27 million units. The rapid growth is thanks to Apple’s entry into the category in the second half of this year. Apple is not only adding a new model or increasing competition in the foldables category; it is converting a segment that was about to decline into the fastest-growing part of the industry.

“

Apple’s entry into the foldable market has done more than reignite growth in a category that was losing momentum. It has fundamentally altered the market’s trajectory. Without Apple, foldable shipments would have declined at a double-digit rate year over year. We forecast Apple will ship more than 17 million foldable iPhones by 2027, capturing roughly 40% of the global foldables market. With an average selling price exceeding $2,550, Apple is expected to generate more than $45.7 billion in value and account for over half of the category’s total value. Perhaps most notably, Apple is positioned to challenge Huawei and Samsung for leadership in markets where they have long dominated, an extraordinary outcome for a product expected to be less than two years into its lifecycle.

What does this mean for consumers?

The days of the cheap smartphone are ending. The average handset now costs roughly $147 more than it did a couple of years ago, and the cheapest models are the ones leaving the shelves fastest. Buyers in price-sensitive markets will feel this loss the most, as the sub-$100 phones many of them relied on are being cut from vendor line-ups. For everyone else it means holding a device for longer and paying more at the point of upgrade. On-device AI is arriving, but the memory it runs on is scarce and expensive, and consumers are covering that cost directly.

What does this mean for vendors?

The next 18 months will separate the vendors who can operate in a structurally more expensive market from those who cannot. Apple, Samsung, and Huawei have the scale and pricing power to turn this challenge to their advantage. Smaller Android brands anchored in the entry tiers face the hardest stretch in the industry’s history, and some will not clear it. The market that emerges on the other side of the crisis, when the memory supply finally stabilizes in 2028, will be smaller in units, larger in value, and far more concentrated at the top. The cheap smartphone era is not pausing. It is over.

Nabila Popal

Nabila Popal - Senior Director, Data & Analytics

Analyst Profile
Nabila Popal is Senor Director with IDC's Data & Analytics team, specializing in Mobile Phones, PC Monitors and other consumer devices.  Ms. Popal is responsible for the global research and quality and timely delivery for her respective technologies, coordinating with regional…
Francisco Jeronimo

Francisco Jeronimo - VP, Data and Analytics, Devices, IDC EMEA

Analyst Profile
Francisco Jeronimo is VP for Data and Analytics at IDC EMEA. Based in London, he leads the research that covers mobile devices, personal computing devices, emerging technologies and the circular economy trends across EMEA. His team delivers data on personal…
Ryan Reith

Ryan Reith - Group Vice President, WW Device Trackers

Analyst Profile
Ryan Reith is the Group Vice President for IDC's Worldwide Device Tracker suite, which includes mobile phones, tablets, wearables, and most recently AR/VR. His teams research focuses on the quantitative aspects of the mobile device industry, including market sizing, forecasting,…

The worldwide enterprise application market grew 11.5% in 2024. It grew 12.1% in 2025. Halfway through 2026, verified earnings across 60 public application software vendors (20 large-cap, 20 mid-cap, and 20 small-cap) put blended growth at roughly 13%, confirmed against the most recently reported quarters. On a market approaching $700 billion, that is real money amounting to tens of billions of dollars in incremental annual spend. It is also, by any honest reading, not the inflection point the AI narrative promised.

That gap is the whole story. If agentic AI were about to remake enterprise software the way cloud remade it a decade and a half ago, the aggregate growth rate should be bending upward. It isn’t bending. It’s creeping. However, therein resides the lie.  An average of a market pulling apart in opposite directions is not a measurement. It’s a number that happens to be both true and useless at the same time.

The reason is not that AI has failed to show up. It is that the aggregate number is the wrong place to look for it. AI capability moves into specific application categories at different speeds, and until it does, it has no reason to move the blended average at all. The businesses waiting for the aggregate number to justify the investment will be years behind the ones who read the dispersion first.

Where the disruption is actually showing up

The real signal was never the aggregate. It is the dispersion beneath it, and 2026 earnings confirm the pattern is holding, not reversing. Atlassian grew 28% in its most recent fiscal quarter. ServiceNow grew 24%. BILL Holdings, an accounts payable and receivable automation platform, grew core revenue 16%. These are workflow-heavy, automation-friendly businesses. This is exactly where agentic AI should add value fastest, and the real numbers agree with the theory.

Customer Service Applications and Human Capital Management tell the opposite story. Sprinklr, a customer experience platform, grew just 6.8% in its most recent quarter, with guidance pointing toward roughly 1% next quarter. LivePerson, sitting in the same customer operations orbit, saw revenue decline 12% year over year. Workday’s total revenue growth held at 13.5% in early 2026, in line with IDC’s full-year 2025 HCM category rate of 13.0%. Not the reacceleration bulls were hoping for.

That is the part worth pausing on. Customer service and HR are arguably the two categories most exposed to AI disruption. Support agents replacing chat queues. AI-driven recruiting replacing manual sourcing and screening. Growth in both is stagnant, not accelerating. The explanation is not that AI failed to show up. Rather, the combination of usage-based, outcome-based, and other agentic pricing models may be cannibalizing legacy per-seat license revenue faster than new AI spend is replacing it. Both could be true, but our research suggests the scale tips toward it being a pricing issue rather than a value issue.

The strongest market cluster is actually not even specifically AI-labeled. It is collaboration software. Atlassian’s 28% and Monday.com’s 22% sit inside this space, and IDC’s 2025 data already showed Enterprise Portals accelerating to nearly 17%.

This is the part akin to the early cloud transition, playing out in real time. The categories with the most surface area for embedded AI agents and assistants moved first, years before analysts had a clean line item to point to. Slack and Teams did not wait for a “collaboration AI” category to exist before AI showed up inside them. Neither will the rest of the market.

What this means now

None of this shows up if the aggregate number is the only thing anyone checks. The categories and the pricing decisions made within them are where the real signal lives, and that holds for tech leaders and software vendors alike.

For tech leaders, the practical version of “don’t trust the aggregate” is to stop asking vendors how their AI roadmap looks and start asking how it will be priced. Before you standardize on any AI feature, ask a specific question: Is this bundled into your current subscription, or will it eventually be metered or split into its own tier? Get the answer in writing. A feature that ships free today can become a paid add-on once your team depends on it, and the time to negotiate a price lock or grandfather clause is before that dependency forms, not after. If a vendor will not commit to how an AI feature will be priced going forward, that ambiguity itself is worth noting and potentially pushing back on in the contract.

The performance of the app category your vendor sits in is a second number worth checking. It tells you something about the pressure that the vendor is under. A vendor in a category where growth has stalled, like customer service software or HCM, is likely facing margin pressure across its customer base, which shapes how it negotiates. First, they may be more willing to discount to retain you. Second, there is a real chance they will get more aggressive about raising prices over time to compensate. Knowing which situation you are in changes how you approach renewal, regardless of how the vendor’s AI roadmap is marketed.

For suppliers, the lesson is direct, and it is the same lesson in reverse. The categories where growth has stalled are not short on AI investment. They are short on pricing models designed to capture the value AI creates rather than giving it away. If your AI features are seeing real adoption and your revenue has not moved, that is not a messaging problem to fix with better marketing. It is a sign your pricing model is absorbing the value your product now creates rather than charging for it. Fix the pricing before you fix the pitch. The vendors in accelerating categories, collaboration and finance automation among them, are not simply better at building AI. They have found ways to attach a price to it.

Both sides are better off treating a vendor’s pricing structure as a more honest signal than its AI messaging, for the same reason the aggregate market number is a less honest signal than the category data underneath it. A vendor confident that its AI is delivering real value usually shows that confidence in how it charges for it. A vendor still bundling AI in for free, indefinitely, may be telling you something about how differentiated that AI actually is, whether or not that is the story in the pitch deck.

The aggregate number will keep lying for as long as it’s the only number anyone checks. Category-level dispersion and the pricing signals underneath it is where the truth actually lives.

Eric Newmark

Eric Newmark - Group Vice President & General Manager of IDC's SaaS, Enterprise Software, CX and Workplace Solutions Division

Analyst Profile
Eric Newmark is Group Vice President & General Manager of IDC’s SaaS, Enterprise Software, CX, and Workplace Solutions Division, which includes several teams of analysts covering SaaS, 18 enterprise application markets, software monetization, business platforms, marketplaces, and services firms focused…

For much of 2025 and into 2026, some engineering teams measured AI success by tokens burned rather than value created, and a handful of companies even built internal leaderboards to track who used AI the most. According to a Washington Post Intelligence report by columnists Dan Gallagher and Kendrick Frankel, that era, nicknamed “tokenmaxxing,” is fading. Airbnb Chief Technology Officer Ahmad Al-Dahle described the shift on LinkedIn, as the Post’s coverage noted: his team tracked throughput and quality instead of token counts, and “usage took care of itself.”

That correction is good news for finance teams tired of chasing a moving number. It does not solve the harder problem underneath it. Most enterprises still cannot say, with confidence, who owns the AI bill, or what it will cost three months from now, a year from now, or three years from now. Token economics, or tokenomics for short, is the discipline meant to answer those questions: managing what AI, and especially agentic AI, actually costs to run, priced per million tokens processed rather than per software seat.

The Bill Nobody Planned For

The Post’s reporting makes the governance gap concrete. It reports that Uber’s chief technology officer told The Information the ride-share company had already exhausted its full-year AI budget by mid-April. It also cites data from a fintech firm showing blended token prices down roughly 50 percent year-over-year, a decline that hasn’t lowered bills so much as fed continuous reasoning loops behind the scenes, so cheaper tokens simply invite more of them.

What IDC’s Own Research Shows

IDC’s research shows the same pattern from the inside, not only from press coverage. In Effective Agent Cost Management: A Practitioner’s Framework for Planning, Sourcing, Implementation, and Operations (IDC #US54644126, June 2026), IDC found that engineering teams typically spot cost anomalies in real time but lack the budget authority to act on them, while finance holds that authority but often can’t detect a problem until a vendor invoice arrives weeks later. “Every dollar of wasted AI agent spend is a governance failure,” says Shari Lava, group vice president, AI, at IDC.

That gap is real, but governance can contain it. IDC’s most recent global cloud AI spending survey, fielded in April 2026 across the Middle East, Türkiye, and Africa, found that 61 percent of organizations exceeded their 2025 cloud AI budget, yet 83 percent of those overruns landed under 15 percent, evidence that early governance narrows overspend even when it can’t prevent it entirely (IDC #META54767326, August 2026). Visibility remains the harder half of the problem. IDC’s Worldwide Technology Buyer and Spending Outlook, based on a May 2026 survey of more than 1,400 IT leaders, found that difficulty budgeting for token- and inference-based pricing is now the single largest barrier organizations report when evaluating AI vendor pricing, ranking above even the unpredictability of usage costs itself (IDC #US54051126, July 2026). The problem is getting worse, not better.

Closing the Ownership Gap

Closing that gap is a shared job. IDC research on CFO-CIO dynamics finds the two roles hold converging influence over technology decisions but distinct, still-unreconciled mandates, and only one in five CEOs have created a separate AI budget line in the first place (IDC #US53393425, July 2025). Most AI spend is still sitting inside a budget category that was never built to isolate it, in the seam between whoever built the model and whoever has to answer for the invoice.

IDC’s governance framework treats that seam as fixable. It calls for naming a single accountable owner per AI workload before the first dollar is spent, building an ownership structure with real authority attached rather than a chart for the board deck, and reviewing token-volume variance as a standing agenda item in the quarterly FP&A cycle instead of waiting for a surprise. The Post’s own recommendations land in a similar place: treat AI spend like any other major budget category, actively monitored, with accountability clearly assigned.

CFOs who want the full framework, including the RACI model and budget governance checklist IDC built for practitioners, can find it in IDC’s ongoing AI Economics and Token Control research, available to IDC Quanta subscribers.

This piece references reporting from “AI tokenomics: Using the technology without busting the budget,” Washington Post/WP Intelligence, August 21, 2026.

Ryan Smith - Content Marketing Director - IDC

Ryan Smith is the Director of Content Marketing at IDC, where he leads brand-level content and social media strategy, aligning research insights with compelling storytelling to engage technology decision-makers. With a background in both IT and marketing, Ryan brings a unique blend of technical understanding and creative strategy to his work. He’s also a seasoned storyteller, speaker, and podcast host who believes the right message, told the right way, can drive both trust and transformation.
Rick Villars

Rick Villars - Group Vice President Worldwide Research

Analyst Profile
Rick is IDC's leading analyst guiding research on the future of the IT Industry. He coordinates all IDC research related to the impact of Cloud and the shift to digital business models across infrastructure, platforms, software, and services. He helps…

近日，2026世界机器人大会在北京举行。本届大会以“人机共生，产需共融”为主题，吸引300余家企业参展，展出超过2000件展品，150余件新品集中发布。值得关注的是，大会首次设立采购日，49家央企以创新联合体形式集中亮相，并围绕12个应用场景展示产业需求，机器人产业的应用导向进一步增强。

需求侧变化正在与产业规模增长形成共振。据工信部数据，2026年1—5月中国机器人规上企业营收突破900亿元，同比增长26.9%；国际数据公司（IDC）数据显示，2025年全球人形机器人出货量约1.8万台，同比暴增近800%。但另一组数据同样值得注意：IDC调研显示，在已开展具身智能机器人试点应用的企业中，大多处于POC阶段，从技术验证到规模部署之间，横亘着稳定性、成本、ROI和运维能力的多重考验。

从本届大会的现场展示来看，机器人正在从单一动作展示转向完整任务执行，从实验室验证走向真实业务场景。Physical AI正在进入从技术能力竞争向产业化能力竞争转变的关键阶段，而本届大会的四个趋势，勾勒出了这条转折路线的清晰轮廓：

  • 从模型竞赛走向真实世界学习，数据成为Physical AI新壁垒
  • 从Demo能力走向规模交付，机器人竞争进入商业化深水区
  • 从单一场景突破走向分层落地，工业制造率先成为规模化主战场
  • 从整机竞争走向生态竞争，产业链加速形成系统化能力

趋势一

数据驱动：从“模型竞赛”走向“真实世界学习”

过去一年，具身智能大模型、VLA模型和世界模型快速发展，推动机器人从感知、理解向复杂任务决策与执行演进。但进入2026年，行业竞争正从“谁的模型更强”进一步转向“谁能让机器人真正干活，并在真实场景中持续学习和进化”。IDC认为，模型是Physical AI的能力基础，而真实世界数据和持续学习能力将成为下一阶段的重要竞争壁垒。

2026世界机器人大会上，世界模型、VLA、仿真平台、机器人训练场及数据采集等技术持续受到关注，虚实结合正成为数据获取的重要路径。通过仿真生成大规模数据并结合真实场景数据，可以降低真实机器人试错成本；机器人运行数据持续回流训练体系，则可形成“感知—训练—执行—反馈”的数据飞轮，推动机器人持续进化。

当前，中国已有14个以上省市布局或建设40余家机器人训练场，成熟训练场数据年产量可达到数百万条，Physical AI的数据基础设施正在加速形成。与此同时，灵巧手、力传感器等硬件能力持续提升，也为机器人获取更加丰富的操作数据提供了基础。

未来，具身智能将进一步形成模型、数据、仿真与机器人本体协同演进的技术体系，企业竞争也将从单纯比较模型能力，转向数据获取、工程化处理和持续学习能力的比拼。

趋势二

交付决胜：从“Demo能力”走向“规模交付”

随着机器人逐步进入真实业务环境，产业竞争的评价标准正在发生变化。本届大会上，多家厂商从单一动作展示转向拆垛、分拣、搬运、配送、清洁、巡检等完整任务验证，机器人正在从展示单项能力转向验证完整任务流程。

机器人产业竞争的分水岭正在从技术Demo能力转向稳定交付能力。

  • 任务能力从单点执行向连续作业升级。机器人需要具备自主感知、任务规划、连续执行及异常恢复能力。
  • 应用部署从POC向规模复制升级。汽车、3C电子、物流仓储等成为重点验证方向，用户对稳定性、部署效率、维护成本及ROI提出更高要求。
  • 评价标准从技术指标向商业指标升级。任务成功率、连续运行时间、单位任务成本和投资回报周期将逐步成为采购决策的重要依据。

IDC用户调研结果显示，已有超过20%的用户从市场关注进入试点探索阶段，但从POC到规模化部署进程正在加速。未来2—3年，稳定运行、快速部署、持续运维和商业ROI将成为机器人规模化落地的关键能力。

趋势三

场景分层：从“单一场景突破”走向“分层落地”

Physical AI不会在所有场景同步实现规模化落地。随着技术成熟度、环境复杂度和商业ROI差异显现，机器人应用正在形成更加清晰的分层路径。当前，具身智能机器人正呈现“服务探索、工业扩张、家庭蓄势”的发展格局。

  • 商用服务进入深入探索阶段。餐饮酒店、零售等场景具有明确的任务需求，但环境复杂程度高于工业场景，对自主导航、环境理解和异常处理能力提出更高要求，具身智能化加速。
  • 工业制造将加速规模化。汽车、3C电子、物流仓储等场景结构化程度较高，任务边界相对清晰，同时存在较强的降本增效需求，是当前技术成熟度与商业价值匹配度较高的应用方向。IDC数据显示，2025年中国工业具身智能机器人市场规模约57.4亿元，其中工业机器人和人形机器人分别贡献约36.2亿元和21.2亿元。
  • 家庭消费市场具备长期潜力。家庭环境高度非结构化，任务类型复杂且长尾，对机器人的泛化能力、安全性、交互能力和成本提出更高要求，其潜在应用场景广泛，长期市场空间更为可观。

未来，机器人应用将沿着高结构化—中结构化—低结构化路径逐步向复杂环境渗透。不同机器人形态也将围绕自身能力优势，与具体场景形成更加明确的匹配关系，推动Physical AI从场景验证走向规模化应用。

趋势四

生态协同：从“整机竞争”走向“系统能力竞争”

随着Physical AI进入产业化阶段，机器人产业链正从单一本体向算力、模型、数据、核心零部件及行业应用全面延伸。本届大会上，国内外产业链企业深度参与，49家央企以创新联合体形式集中亮相，产业竞争正从单一技术供给走向产业链协同与场景共创。

随着Physical AI正在推动机器人产业从整机竞争走向系统能力竞争，企业竞争边界将进一步扩展至软硬件协同与产业生态。

  • 产业链协同持续深化。芯片、传感器、减速器、伺服系统、灵巧手等核心部件与机器人本体加速协同，供应链能力持续提升。
  • 模型与本体加速融合。大模型企业、机器人厂商及科研机构围绕VLA、世界模型、运动控制等开展合作，推动AI能力进一步进入真实机器人和实际应用。
  • 产业支撑体系加速完善。中国各地方已落地数十个具身智能相关创新中心，覆盖技术攻关、中试验证、数据训练及场景落地等环节，产业协同基础持续增强。
  • 国际合作持续深化。大会吸引30家国际支持机构参与，覆盖科研、工程、产业及投资等领域，推动全球机器人技术交流、产业协同与创新成果共享。

未来，企业竞争将从单一产品能力进一步转向软硬件协同、数据闭环、产业链整合和行业交付能力。具备完整生态组织能力的企业，将更有机会推动机器人从单点应用走向规模化商业部署。

IDC未来展望

WRC 2026释放出的信号越来越明确：具身智能机器人产业正在从技术展示走向产业价值验证。

未来机器人企业的核心竞争力，将不只是技术先进性，更在于能否把技术转化为稳定、可复制、具备商业价值的产品与解决方案。

Physical AI正在进入产业化深水区，机器人行业的下一阶段竞争，也将更加聚焦真实需求与实际价值。

IDC相关研究

IDC长期追踪机器人与Physical AI产业发展，形成从市场追踪、技术评估到用户洞察的持续性研究体系，并在工业制造、商用服务、特种作业及家庭消费等场景方向均有深入研究布局。如需了解更多IDC在机器人领域的研究内容或获取定制化数据分析，欢迎与IDC中国分析师团队联系。

进一步沟通

机器人产业正站在从技术验证迈向规模应用的关键节点。若您关注Physical AI落地路径、具身智能场景价值评估或产业链竞争格局，欢迎与IDC中国分析师团队深入交流。我们将结合WRC 2026现场一手观察与持续研究积累，为您提供定制化洞察与数据支撑。请点击此处进行联系。

Lily Li

Lily Li - Research Manager

Analyst Profile
Lily is the Research Manager for China Robotics and Embodied Intelligence, specializing in market research on embodied intelligent robots. She has long focused on the development trends of China’s embodied intelligence robotics industry, systematically studying the evolution of robot hardware,…

生成式AI正以前所未有的速度渗透到企业的核心业务中——无论是研发、运营还是客户服务，AI都在重塑我们的工作方式。但你有没有注意到，随着AI系统架构越来越复杂，安全风险也在同步升级？

过去我们担心的只是模型越狱、敏感信息泄露这些“老问题”，现在，提示注入、知识库投毒、工具滥用、智能体越权、多智能体协同攻击等新型威胁层出不穷。企业的安全治理对象，早已从单一模型扩展到整个生成式AI应用系统，安全边界变得模糊且动态。

安全评估的方式也在快速演进——从人工测试到自动化红队、从静态规则到动态风险评分，再到如今覆盖RAG、智能体和系统级的持续安全监测。企业管理者们越来越关心：我的AI系统到底有多“可信”？上线前后如何持续掌控风险？

IDC认为，生成式AI安全评估平台正从“单点测试”走向“系统级守护”，未来将深度融入AI开发、部署、运营和治理的每一个环节，实现从一次性检测到全流程、持续的安全验证。只有这样，企业才能真正让AI成为安全、可信的生产力，而不是新的风险源。

国际数据公司（IDC）特别针对中国市场发布了《IDC MarketScape：中国生成式AI的安全评估平台厂商评估，2026》（Doc# CHC54110426，2026年8月）研究报告，IDC针对入选的提供商的能力和战略进行综合评估，评估标准包括收入规模、产品技术能力、行业和客户拓展、生态系统建设以及未来发展战略等方面，综合分析并确定各厂商在IDC MarketScape中的相对位置，并探讨厂商的特点和战略，总结发展趋势，为企业级客户选择生成式AI的安全评估平台能力提供商提供参考。

报告揭示了一个关键信号：IDC认为，头部厂商的竞争焦点已从“单点测试工具”转向“系统性安全治理平台” ——这一转变直接决定了企业该选择什么样的合作伙伴，以及该用什么样的标准去判断谁更值得信任。报告的完整评估结果与厂商定位分析，可向IDC进一步了解。

未来，企业关注的不再只是模型是否能够抵御越狱攻击，而是更加关注整个AI应用系统在实际业务场景中的安全表现，包括知识库安全、智能体运行安全、工具调用安全以及业务流程安全等多个方面。与此同时，评估能力也将逐步融入企业AI开发、部署、运营及治理全过程，实现由”一次性安全测试”向”持续安全验证”转型。

生成式AI的安全评估平台未来演进方向

  • 评估对象持续扩展。随着生成式AI技术持续向智能体、多智能体协同及具身智能等方向发展，企业AI系统的复杂度和应用边界将不断拓展。未来，生成式AI的安全评估平台将逐步突破单一模型评测范畴，形成覆盖基础模型、AI应用、智能体、具身智能及相关生态组件的综合评估体系，持续提升对新型AI应用场景的风险识别与安全验证能力。
  • 评估体系更加完善。未来安全评估将更加注重建立统一、标准化的评估体系，在覆盖风险类型、评估维度及评价指标等方面持续完善，提高不同产品之间评估结果的可比性，为企业提供更加客观、可信的评估依据。
  • 智能化能力持续增强。人工智能技术将进一步融入安全评估过程，自动化红队测试、智能风险分析、攻击样本自动生成及评估结果智能分析等能力将不断成熟，进一步提升风险发现效率和评估质量。
  • 持续评估能力成为核心竞争力。安全评估将逐步融入企业AI研发、测试、部署及运营全过程，形成覆盖模型上线前验证、运行中监测以及版本迭代评估的持续安全验证体系，为企业提供动态、实时的风险管理能力。
  • 平台化能力不断提升。未来生成式AI的安全评估平台将进一步与AI开发平台、安全运营平台及AI治理平台深度融合，由独立评估工具逐步发展为企业AI安全治理的重要基础设施，为企业构建可信生成式AI体系提供持续支撑。

IDC给技术买家的建议

  • 建议结合企业AI应用现状制定评估需求。不同行业、不同规模企业的生成式AI建设阶段存在较大差异，安全评估需求也各不相同。技术买家应结合自身AI应用场景、业务规模及安全治理目标，明确评估范围和重点，避免单纯关注评测数量或功能覆盖，而应重点考察产品是否能够满足企业实际业务需求。
  • 建议重点关注评估体系的科学性和完整性。安全评估平台的核心价值不仅体现在风险发现能力，更体现在评估体系的完整性和结果的可信度。技术买家应关注产品是否建立了完善的评估指标体系、风险分类体系及评分机制，评估结果是否具有可解释性和可追溯性，是否能够为企业风险治理提供有效依据。
  • 建议关注产品的持续演进能力。生成式AI技术快速发展，新的应用形态和安全风险不断出现，安全评估平台需要持续更新评测能力和风险场景。技术买家应重点考察厂商在技术研发、规则更新、测试场景维护及产品迭代方面的能力，确保产品能够适应未来AI技术的发展需求。
  • 建议关注平台的开放性与生态兼容能力。企业AI应用通常涉及多个基础模型、开发框架及业务平台，安全评估平台应具备良好的开放能力，支持多模型、多场景及多种部署方式，并能够与企业现有AI开发平台、安全运营平台及治理体系实现有效集成，降低后续建设和运维成本。
  • 建议将安全评估纳入AI治理体系建设。生成式AI的安全评估不应作为模型上线前的一次性测试，而应贯穿AI研发、部署、运营及持续优化全过程。技术买家应推动安全评估与AI开发流程、安全运营流程及风险治理机制相结合，逐步建立持续、动态的生成式AI的安全评估机制，提升企业整体AI安全治理能力。
  • 建议结合实际业务场景开展产品验证。生成式AI的安全评估平台的能力最终需要在真实业务环境中得到验证。技术买家在产品选型过程中，建议结合自身AI应用场景、典型业务流程及安全需求开展概念验证（PoC），重点考察产品在风险识别准确性、评估覆盖范围、自动化水平及评估结果稳定性等方面的实际表现，避免仅依据产品演示或功能清单进行判断。
  • 建议综合评估厂商的持续服务与运营能力。生成式AI的安全评估是一项持续性的工作，产品能力需要随着AI技术演进和风险变化不断更新。技术买家除关注产品功能外，还应重点评估厂商的技术研发能力、评测体系更新能力、行业实践经验、专业服务能力及持续运营支持能力，优先选择能够提供长期技术服务和持续能力演进的合作伙伴，为企业生成式AI安全治理提供长期支撑。

IDC中国网络安全分析师陈佳表示，生成式AI安全评估平台正由早期合规检测工具向企业AI治理的基础安全能力平台演进，企业关注点从“可用”转向“安全、可靠、可控”，平台能力覆盖模型、数据、内容、智能体安全及持续运营管理。随着行业标准完善，评估指标、测试集和自动化能力成为厂商差异化关键，未来相关能力将平台化融合，推动生成式AI从“可用”迈向“可信”。

进一步交流

生成式AI安全治理进入关键阶段。IDC持续关注生成式AI安全评估平台技术演进与市场实践，助力企业识别模型风险、完善治理体系、构建可信AI能力。欢迎联系IDC，获取行业洞察与定制化评估服务，共同探索生成式AI安全落地新路径。

Jackson Chen

Jackson Chen - Research Analyst

Analyst Profile
Jackson Chen is a research analyst for IDC China's Emerging Technology Research department, whose focus is on research of the business sector of the China data security market. Jackson is also responsible for providing relevant security research and consulting services…

8月19日，2026世界机器人大会（WRC 2026）在北京亦庄开幕。今年大会以“人机共生，产需共融”为主题，并首次设置采购日，49家央企携12个应用场景集中亮相，进一步强化产业供需对接。

从大会主题和组织形式来看，机器人产业正从技术展示和单点突破，逐步转向产品交付、场景落地与产业协同。随着机器人本体、AI模型、核心零部件和数据基础设施持续发展，产业竞争也正从单一产品能力转向软硬件、数据、模型与应用场景的综合能力竞争，WRC 2026成为观察这一变化的重要窗口。

从技术展示走向需求验证，产业化进入关键阶段

近两年，人形机器人、四足机器人和具身智能等方向持续获得资本和产业关注，运动控制、环境感知、智能决策和操作能力不断提升。国际数据公司（IDC）用户调研显示，中国具身智能机器人正从展示型创新逐步进入价值验证与规模复制阶段，超过20%的用户已从关注进入试点探索。

今年WRC首次设置采购日，进一步强化产业需求导向。与此同时，政策也持续推动机器人进入真实应用场景，通过实景实训专项行动、央企机器人创新联合体等方式，加快B端需求释放。

随着机器人进入真实场景，商业化评价标准也将从功能和技术指标进一步延伸至任务成功率、连续运行能力、故障率、部署成本和ROI等指标。需求能否转化为订单、POC能否进入批量部署，将成为衡量产业化成熟度的重要指标。

Physical AI加速重塑机器人智能化路径

机器人从技术展示走向场景落地，也推动Physical AI进入产业视野。Physical AI代表AI从数字世界走向物理世界的技术与产业范式，其核心是让AI具备对真实物理环境进行自主感知、理解、决策和行动的能力；具身智能机器人则是其重要的物理载体。

IDC 2026年全球CEO调查显示，35.2%的受访CEO将Physical AI列为未来12—24个月重点关注的新技术投资方向。在WRC 2026，世界模型、视觉-语言-动作模型（VLA）等技术正加速与机器人本体、运动控制和数据系统融合。

从产业化角度看，关键不只是机器人“有多智能”，更在于模型、数据、本体与应用场景能否形成有效闭环，并在真实环境中稳定运行、持续迭代。这将成为Physical AI走向规模化应用的重要验证。

六大产业环节共同构成Physical AI产业基础

Physical AI的发展正在突破单一机器人本体或AI模型的边界，逐步形成由算力、模型、数据、本体、核心零部件和应用共同支撑的产业体系。

IDC最新发布的《MarketGlance: 中国具身智能机器人，3Q26 》报告，将当前机器人与具身智能产业划分为六大细分市场：算力基础设施、模型算法、数据引擎、机器人本体、核心零部件与基础设施，以及行业应用智能体。六大环节共同构成具身智能机器人从自主感知、理解、学习到行动的完整产业链。

这一产业格局也在WRC 2026得到集中呈现。智合、智创、智造、智趣四大展馆覆盖从底层技术、核心零部件、机器人本体到行业应用的多个环节，呈现出机器人产业由单一整机竞争向全产业链协同和生态竞争演进的趋势。

WRC 2026值得关注的四个趋势

基于对产业链和技术发展的持续跟踪，IDC认为，今年WRC有四个趋势值得重点关注。

竞争从“技术领先”转向“交付领先”。

机器人产业真正的分水岭正在从Demo能力转向稳定运行和规模交付能力，企业能否完成从技术验证到批量部署的跨越，将成为下一阶段竞争的重要差异化因素。

机器人从“能力展示”走向“价值验证”

市场关注点正从机器人形态和单项技术能力，进一步转向任务完成能力、运行稳定性、应用效率和经济性。随着人形、四足及其他具身智能机器人加速进入真实场景，能否稳定完成实际任务并形成明确的应用价值，将成为下一阶段商业化竞争的关键。

Physical AI从“模型竞赛”走向“真实世界学习竞赛”。

模型能力只是起点，数据获取、场景泛化和持续迭代能力将成为新的技术壁垒。能够通过真实场景持续积累数据并反哺模型的企业，有望形成更强的智能化能力闭环。

产业竞争从“整机竞争”走向“生态竞争”。

随着模型、数据、本体、核心零部件与应用之间的联系不断增强，企业的竞争边界正在从单一产品延伸至产业链协同能力，生态构建能力的重要性持续提升。

IDC观点：机器人产业进入能力与商业化并重阶段

WRC 2026释放的信号表明，机器人产业正从技术能力竞争转向产业化能力竞争。随着本体、模型和核心零部件持续成熟，稳定运行、规模交付、数据积累和软硬件协同能力将成为下一阶段竞争关键。

未来2-3年将是具身智能机器人从技术验证走向规模化应用的重要窗口期，不同场景将呈现分层落地特征：服务场景率先开展探索应用，工业场景将成为具身智能机器人规模化落地的重点方向，家庭消费市场则具备长期发展潜力。与此同时，Physical AI将推动产业竞争从单一硬件能力向“模型+数据+本体+应用”的系统能力演进。

WRC 2026的开幕，是观察这一变化的重要节点。IDC分析师团队将在现场持续关注应用落地、Physical AI技术进展及产业链竞争格局，进一步复盘WRC 2026，敬请关注IDC中国后续研究与现场探展内容。

进一步交流：

机器人产业正站在从技术验证迈向规模应用的关键节点。若您关注Physical AI落地路径、具身智能场景价值或产业链竞争格局，欢迎与IDC中国分析师团队深入交流。我们将结合WRC 2026现场观察与持续研究，为您提供定制化洞察与数据支撑。点击此处联系我们。

Lily Li

Lily Li - Research Manager

Analyst Profile
Lily is the Research Manager for China Robotics and Embodied Intelligence, specializing in market research on embodied intelligent robots. She has long focused on the development trends of China’s embodied intelligence robotics industry, systematically studying the evolution of robot hardware,…

At AMD Advancing AI 2026, held on July 23, 2026, at the Moscone Center in San Francisco, California, AMD presented a broad AI strategy that spans cloud infrastructure, enterprise AI, client computing, robotics, and embedded systems. The event brought together developers, customers, ecosystem partners, and enterprise leaders to discuss how AMD’s end-to-end AI portfolio is evolving from silicon through software.

Financial Momentum Supports the AI Strategy

The announcements came shortly before AMD reported its second-quarter 2026 financial results. The company delivered record revenue of $11.5 billion, representing 50% year-over-year growth, while datacenter revenue more than doubled compared to the prior year. AMD attributed this growth to accelerating demand for EPYC processors, expanding Instinct deployments, and increasing AI infrastructure investments across the market.

A Vision Beyond Silicon

While hardware announcements often dominate AI events, the broader message from Advancing AI 2026 was AMD’s effort to create a unified AI development experience. The company’s strategy aims to enable developers to build AI applications on local systems, optimize them with a common software framework, and deploy them consistently across enterprise, cloud, edge, and robotics environments using the same foundational platform.

Building an Open AI Software Foundation

A key pillar of AMD’s strategy is its continued investment in open software. AMD reported that more than three million Hugging Face models run out of the box on AMD platforms, while many of the industry’s leading open-source AI projects now offer native AMD support. The company also highlighted significant growth in open-source contributions and ecosystem partnerships as part of its software expansion efforts.

ROCm Evolves from Platform to Development Ecosystem

AMD’s ROCm platform has evolved from a GPU programming environment into a broader AI software stack that supports frameworks, libraries, developer tools, and optimization capabilities. AMD positions ROCm as the common software layer that enables portability across its AI hardware portfolio, from client systems through hyperscale infrastructure.

Introducing ROCm.ai

Among the notable software announcements was ROCm.ai, an AI-driven development platform designed to help developers build, optimize, and deploy GPU applications more efficiently. ROCm.ai enables popular coding assistants such as Claude, Codex, Cursor, and Gemini to understand AMD hardware architectures and ROCm workflows natively, helping reduce the learning curve associated with GPU optimization.

AI-Assisted Performance Optimization

AMD also introduced capabilities that leverage AI to improve application performance. Through technologies such as Hyperloom, ROCm.ai can establish performance baselines, profile workloads, identify bottlenecks, tune serving configurations, generate optimized GPU kernels, and validate performance improvements. This represents a shift toward AI-assisted software engineering where optimization becomes increasingly automated.

Extending AI Development to the Desktop

AMD’s vision extends beyond the datacenter. With the Ryzen AI Halo developer platform and upcoming systems powered by Ryzen AI Max PRO 400 Series processors, AMD aims to make advanced AI development accessible on desktop-class hardware. These systems provide developers with local environments capable of running increasingly sophisticated AI workloads while maintaining compatibility with larger AMD compute environments.

Ryzen AI Halo developer platform, source: AMD, 2026

Developing Locally, Scaling Globally

A notable aspect of AMD’s approach is software consistency across deployment environments. AMD presented ROCm.ai as supporting systems that range from AI PCs and workstations to rack-scale AI infrastructure and datacenters. This approach creates a workflow where models can be developed and tested on local hardware before being deployed at cloud scale using the same software foundation.

Enterprise AI and Hybrid Deployment

AMD also highlighted a collaboration with Cisco designed to combine AMD AI inference platforms, including Ryzen AI Halo systems, with Cisco networking, security, observability, and management capabilities. The goal is to help enterprises deploy, govern, and manage hybrid and local agentic AI deployments while maintaining operational consistency across environments.

Extending AI into Robotics and Physical Systems

Another major theme of the event was AMD’s expansion into Physical AI. As AI moves beyond digital environments into machines that perceive, reason, and act in the physical world, AMD is positioning its technology portfolio to address robotics, industrial automation, healthcare, and autonomous systems. AMD estimates that the Physical AI silicon market could reach $200 billion by 2035.

Introducing the Kria AI Portfolio

To address this opportunity, AMD introduced the Kria AI solutions portfolio, extending its presence from traditional robotics control systems into higher-level robot intelligence. AMD describes this as bringing together perception, reasoning, decision-making, and control capabilities on a unified platform designed for demanding robotic workloads.

Ryzen AI Embedded X100 Series

At the foundation of the portfolio is the new Ryzen AI Embedded X100 Series, which combines CPU, GPU, and NPU resources within a unified memory architecture. AMD positions these processors as platforms capable of supporting the simultaneous requirements of perception, reasoning, planning, orchestration, and real-time control that characterize modern Physical AI systems.

Ryzen AI Embedded X100 Series, source: AMD, 2026

The Kria AI Robotics Developer Platform

AMD also introduced the Kria AI Robotics Developer Platform, described as an open, turnkey robotics development environment that integrates CPU, GPU, NPU, and FPGA compute. The platform is intended to reduce development complexity while helping robotics developers move more quickly from concept and prototype to production deployment.

Kria AI Robotics Developer Platform, source: AMD, 2026

Bringing the Cloud Software Stack to the Edge

One strategically important aspect of AMD’s robotics announcements is software continuity. The new AMD Robotics Software Suite is built on ROCm, allowing developers to leverage familiar frameworks, tools, and workflows already used in cloud AI environments. AMD explicitly describes ROCm as both cloud-proven and embedded-ready, creating a common software layer between datacenter AI and Physical AI deployments.

Reducing Vendor Lock-In Through Open Standards

AMD repeatedly emphasized openness throughout its software and robotics announcements. The company highlighted support for open standards, familiar x86 development workflows, industry partnerships, and open-source technologies. In robotics specifically, AMD identified avoidance of vendor lock-in as a key priority among developers and positioned its software stack as an alternative built around portability and ecosystem flexibility.

Analyst Opinion and Conclusion

AMD’s Advancing AI 2026 announcements reflect a strategy focused on developer productivity, software consistency, and open ecosystem adoption. Rather than treating desktops, datacenters, edge systems, and robotics as separate markets, AMD is building a continuum in which applications can be developed locally, optimized through AI-assisted software tools, deployed at cloud scale, and extended into physical systems using a common architecture. This approach addresses a persistent industry challenge: the fragmentation between experimentation, production deployment, and edge execution.

Technical vision alone will not guarantee adoption, however. AMD continues to compete against an incumbent ecosystem built on years of developer investment, mature tooling, and deep integration with commercial AI applications, and shifting those established development priorities typically requires clear performance, cost, or business benefits. Pricing and total cost of ownership will matter as much as raw compute performance, particularly as enterprise buyers weigh software readiness, operational simplicity, and deployment timelines across the full lifecycle.

Success will likely depend less on hardware specifications than on ecosystem momentum: winning ISV certifications, application optimizations, framework support, and robotics middleware integrations may prove as important as introducing faster processors. The broader significance of Advancing AI 2026 is not that AMD is entering new AI segments, but that it is attempting to connect desktop, cloud, and physical AI through a common development model. Whether AMD becomes an increasingly influential platform provider across that continuum will depend on its ability to sustain software adoption and partner engagement alongside its hardware roadmap.

Mohamed Hefny

Mohamed Hefny - Senior Program Manager, Data and Analytics

Analyst Profile
Mohamed Hefny leads market research in EMEA on professional workstation PCs and solutions. He also reports on professional computing semiconductors, processors, and accelerators (CPUs and GPUs), as well as breakthroughs and trends related to the market. In addition, Mohamed is…
Malini Paul

Malini Paul - Senior Research Manager, Personal Computing Devices

Analyst Profile
Malini Paul is a research manager of Client Computing Devices team in Europe, with over 10 years of experience in the ICT industry. Currently based in London, she leads the Personal Client Computing Devices research program for Western Europe. She…