Walk through a laptop aisle in 2026 and the Copilot+ PC branding is highlight for most Windows laptops. From Microsoft’s own surface to other PC makers like Samsung, HP, and Dell, you can find notebooks that carry this badge to convey that they are AI-ready. At a glance, the name sounds like it refers to a computer with a better version of the Copilot chatbot, which only explains a small part of it.

A Copilot+ PC is a Windows 11 computer that meets Microsoft’s hardware standard for advanced on-device AI features like a compatible processor with a dedicated NPU. You also need a certain amount of RAM and storage, all of which brings access to Windows features such as Recall, Click to Do, and much more. Many of these experiences use the NPU to process information locally, reducing their reliance on cloud servers and helping them run more efficiently in the background.

The badge has expanded considerably since the first Copilot+ laptops arrived in June 2024. Snapdragon X processors were initially the only option. Current models can also use qualifying AMD Ryzen AI and Intel Core Ultra chips, giving buyers a choice between Arm and conventional x86 Windows systems. So here’s everything you need to know.

What is a Copilot+ PC?

Copilot+ is Microsoft’s certification for a class of Windows 11 AI PCs. A qualifying computer combines a sufficiently powerful NPU with Microsoft’s minimum memory and storage requirements. The Copilot app itself does not require this hardware. A regular Windows 11 PC can still access Microsoft Copilot since it relies on an internet connection. On the other hand, Copilot+ systems gain a separate set of Windows features designed around local AI processing.

It is also worth noting that buying a Copilot+ PC also does not automatically include a premium Copilot or Microsoft 365 subscription. Most built-in Windows experiences are part of the operating system, however, certain actions and connected services may require an account, internet access, or an additional subscription.

Copilot+ PC hardware requirements

As of right now, Microsoft has currently named these processor familiar as compatible:

• AMD Ryzen AI 300 and 400 series
• Intel Core Ultra 200 and 300 series
• Qualcomm Snapdragon X series

A processor carrying one of those broader family names does not guarantee that every configuration supports every feature. Buyers should still look for the actual Copilot+ PC badge and check the manufacturer’s specifications.

What are TOPS, and why do you need an NPU?

Copilot+ PCs are all about AI. However, this was one of the biggest problems with previous-gen AI PCs. They had NPUs, yes, but little to do with them.The NPU is a dedicated part of the processor built to handle AI workloads efficiently. CPUs and GPUs can also run AI models, but the NPU is designed for sustained jobs such as background effects, image analysis, speech processing, and semantic search without placing the same load on the main processor or graphics hardware.

Microsoft measures the minimum NPU requirement in TOPS, short for trillion operations per second. A Copilot+ system needs at least 40 TOPS of NPU performance. This number only describes one part of the computer. It does not tell you how fast the CPU is, or how well it can run games. So two systems carrying the same Copilot+ badge can deliver very different everyday performance.

AI PC versus Copilot+ PC

“AI PC” is a flexible industry term. Manufacturers commonly use it for computers with an NPU or other hardware intended to accelerate AI tasks. Copilot+ PC is Microsoft’s more tightly defined category. Every Copilot+ computer is an AI PC, while many systems marketed as AI PCs fall below Microsoft’s 40TOPS requirement or lack access to the complete Copilot+ feature set. A great example of this is an older Intel Core Ultra laptop, which may advertise an NPU and AI features, while missing the Copilot+ badge because its NPU does not meet Microsoft’s threshold.

Are all Copilot+ PCs Arm computers?

No. The first Copilot+ systems used Qualcomm’s Arm-based Snapdragon X processors, which made the category look closely tied to Windows on Arm. Intel and AMD have since added qualifying x86 chips. But compatibility isn’t consistent. While Snapdragon models can offer strong responsiveness and battery efficiency, Windows 11 uses Microsoft’s Prism emulator to run many x86 and x64 applications that do not have native Arm versions.

What AI features do Copilot+ PCs include?

Microsoft’s Copilot+ feature list has grown considerably since launch. Availability can vary by processor, region, language, account type, and Windows update. So here are the highlights:

Recall

Recall creates an optional, searchable timeline using snapshots of activity on your screen. You can describe a document, webpage, image, or app you remember seeing and search through the saved timeline to locate it. Microsoft continues to label Recall as a preview.

Click to Do

Click to Do analyzes selected text and images on the screen, then offers relevant actions. Depending on the content, it may let you copy text, summarize or rewrite it, search the web, remove an image background, blur a background, or open the selection in another app. Some actions require a subscription.

Improved Windows Search

Improved Windows Search uses semantic indexing so you can find supported files and images through natural descriptions. Searching for “team at the conference” can work without knowing the exact filename. It appears in File Explorer, the Windows search box, and supported Settings searches.

Agent in Settings

Agent in Settings lets you describe a Windows problem or setting in ordinary language. It can surface the relevant option and, for supported changes, help apply it. Microsoft expanded the feature’s language support in 2026.

Live Captions with translation

Windows 11 already supports Live Captions. Copilot+ systems add local translation from more than 40 languages into English and from supported languages into Simplified Chinese.

Creative and accessibility tools

Copilot+ PCs also support AI tools across Paint, Photos, Snipping Tool, Voice Access, Narrator, and Windows Studio Effects. These include Cocreator, Restyle Image, Image Creator, Super Resolution, Perfect Screenshot, flexible voice commands, richer image descriptions, background blur, eye contact, automatic framing, and voice focus.

Some tools still have processor restrictions. Automatic Super Resolution, Paint Generative Fill, and Photos Relight currently list Snapdragon X requirements on Microsoft’s feature page.

Is Recall private?

Recall attracted heavy criticism when Microsoft first announced it, which led the company to delay its broad release and redesign the security model. This concern was also reflected in our own experience with the feature. The current version is opt-in and requires Windows Hello Enhanced Sign-in Security with biometric authentication. Snapshots are encrypted, stored locally, and tied to the individual Windows profile. Microsoft says it cannot access them, and other apps cannot retrieve the Recall database.

Sensitive-information filtering is enabled by default to help avoid saving passwords, payment details, and identification numbers. Users can pause snapshots, delete them, limit storage, and exclude specific apps or websites. Recall can also be removed as an optional Windows component.

Storage is another consideration. Microsoft says Recall needs at least 50GB of free space to operate. A 256GB system allocates 25GB to snapshots by default, which Microsoft estimates can retain about three months of activity. Older snapshots are deleted once the allocation fills.

Are Copilot+ PCs faster and more efficient?

Microsoft markets Copilot+ systems as its fastest and longest-lasting Windows PCs. Its commissioned testing has claimed up to 22 hours of local video playback and 15 hours of web browsing on selected devices. Those figures vary with the processor, display, battery size, workload, settings, and manufacturer. But these figures are often seen in marketing material for recent high-end laptops.

The Copilot+ badge itself is not a performance ranking. Qualcomm systems often prioritize efficiency, AMD offers great integrated graphics and strong multi-core performance, while new Intel Core Ultra processors can either focus on battery life or strong raw horsepower. So traditional specifications are still important.

Should you buy a Copilot+ PC?

A Copilot+ PC branding alone isn’t enough to justify a purchase, but if you’re looking for a high-end or upper mid-range model, chances are you’re going to get one anyway. Only the entry level models rely on processors based on older architecture that don’t get the shiny new NPUs. Another limitation currently is with RAM capacities. Owing to the RAMmageddon, 8GB machines are returning, so if you’re on a tight budget, Copilot+ PCs might be a little too expensive.

Regardless, you’re not missing out on much–at least for now that is. AI tools like Recall or Click to Do can be handy, but they won’t radically change the way you use or experience your Windows PC. I have a Copilot+ PC of my own, but I hardly ever found myself using these features.

A decade ago, then-Uber CEO Travis Kalanick said he saw autonomous vehicles as an existential threat to the ride-hail company’s business model.

“What would happen if we weren’t a part of that future? If we weren’t part of the autonomy thing? Then the future passes us by,” Kalanick told Business Insider.

In the years since, Uber has settled on a strategy that, rather than see it build and operate its own self-driving cars, puts it on track to become the place where riders can get connected with any ride, driven by a human or robot. “We think there are going to be many AV players around the world, and we want to be the go-to commercial platform for all of them,” now-CEO Dara Khosrowshahi told investors in 2024. Since then, the company has signed agreements with more than 25 major robotaxi players, with driverless vehicles from Waymo, Nuro, Baidu, and Volkswagen’s MOIA either available or soon to be available on the Uber app in several global cities.

Now, according to documents viewed by WIRED and another obtained through a public records request, Uber’s lobbyists are pushing to build that strategy into law. The company’s representatives have pressed lawmakers to deploy autonomous vehicles on what it calls “hybrid networks,” where human drivers work alongside robots as the new tech grows.

In New Jersey, a lobbyist representing Uber took the strategy a step further, circulating legislative language that would, for a period of three years, require any platform offering driverless ride-hailing services to have human drivers serve 85 percent of its rides.

The language would likely prevent self-driving vehicle developers, including Waymo, Zoox, and Tesla, from operating their own ride-hail apps in the state—effectively forcing them onto another ride-hail app if they hope to enter the market and limiting competition for Uber, the country’s reigning ride-hail leader.

A representative for Uber pitched a version of the proposal to New Jersey state senator Andrew Zwicker, according to his chief of staff, Ayla Rios. Zwicker is the sponsor of a bill currently being considered by the state legislature that would establish New Jersey’s first set of rules governing self-driving cars on public roads. The Uber lobbyists’ proposed language restricting standalone robotaxi-hailing apps is not currently part of the bill, which could come up for a vote this fall.

The New Jersey bill is the first proposed in the nation that would limit the operation of Tesla’s robotaxis, because it requires AV developers to use multiple sensors to power its software, rather than just cameras, as Tesla’s technology does. It would also require vehicles to be operated in emergencies using steering wheels and brake pedals, which purpose-built robotaxis like those from Zoox do not have.

In Washington, DC, where autonomous vehicle developers, including Waymo, are engaged in a pitched, months-long battle to allow robotaxi services to operate in the district, Uber representatives also sought to ensure that “hybrid networks” would be the future of ride-hail.

A bill introduced by city council member Charles Allen in April would allow driverless services on DC’s public roads under certain conditions. In an email sent more than a week before the introduction of the legislation and obtained by WIRED through a public records request, Uber lobbyist LáVita Gardner thanked an Allen staffer for committing to allowing ride-hail companies like Uber to participate in the district’s autonomous vehicle program. “Allowing for hybrid networks will be critical for a smooth transition that supports both technology and human drivers,” Gardner wrote. (The DC bill will be the subject of a hearing on Monday, and has not yet come up for a vote.)

Strands is the NYT’s latest word game after the likes of Wordle, Spelling Bee and Connections – and it’s great fun. It can be difficult, though, so read on for my Strands hints.

Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc’s Wordle today page for the original viral word game.

Highly anticipated: CUDA has become so embedded in high-performance computing that most developers treat it as inseparable from Nvidia hardware. A small London startup is trying to change that by making CUDA code run across different chips without forcing developers to start over. Spectral Compute has built a compiler called SCALE that serves as a drop-in replacement for Nvidia’s NVCC, letting developers run existing CUDA code on other hardware, including AMD GPUs, without rewriting it.

Spectral Compute was founded in 2018 by four engineers with about 60 years of combined HPC optimization experience. The founders say the effort grew out of frustration: while working at an AI firm, they grew tired of the cost of Nvidia GPUs and the poor performance of alternative compilers, which pushed them to build their own solution using LLVM and Clang.

Unlike tools that translate CUDA into another language or operate on already compiled binaries, SCALE works as a compiler in its own right, recompiling CUDA directly for the target hardware. The model follows the way CPU compilers work, where code can run on different architectures and performance differences mostly come from the hardware, not the compiler.

Spectral is working from the assumption that CUDA is here to stay, noting that it accounts for about 80% of HPC code in use today.

“We take the approach that’s industry-standard for CPUs, but apply it to GPUs,” Giulio Malitesta, head of growth at Spectral, told HPCwire. He added that it’s “the same approach that enables C++ to run, for example, on AMD and ARM CPUs, where nobody expects a performance gap that isn’t directly caused by differences in the underlying hardware.”

Spectral is working from the assumption that CUDA is here to stay, noting that it accounts for about 80% of HPC code in use today. “CUDA is basically the de-facto standard of HPC,” Malitesta said. “We need to accept that as a fact and just do the work as compiler engineers to make it available on different platforms that are not necessarily Nvidia, but also improve on Nvidia GPUs.”

Several other tools also aim to make CUDA portable, but each has notable limitations. AMD’s HIPIFY converts CUDA code into C++ for its ROCm platform, but it doesn’t fully leverage low-level features such as PTX. Intel’s SYCLomatic migrates about 90% of the code, leaving the remaining 10% for manual cleanup. Tools like ZLUDA work at the binary level, which can hurt performance.

Spectral argues its method avoids those tradeoffs. By recompiling from source and checking results against NVCC outputs, the company says it can preserve accuracy while improving performance. Benchmarks published by Spectral show that SCALE can significantly outperform HIPIFY-based approaches on AMD GPUs, with gains in some cases approaching six times.

So far, the company has focused on AMD hardware but is working toward supporting other AI accelerators, though it hasn’t named them. It also continues to support Nvidia GPUs, where it believes there is still room to improve performance through better compilation.

The broader CUDA ecosystem adds another layer of complexity. There are hundreds of specialized libraries, including cuDNN, cuTENSOR and cuDF, that many applications depend on. Spectral is working to expand support for those, and it plans to roll out PyTorch compatibility to better integrate with common AI workflows.

Even as it works to make CUDA more portable, Spectral says it is not trying to compete directly with Nvidia. The company joined Nvidia’s Inception program in June and says it is working across the industry. “We’re on the good side of Nvidia and we’re on a good side with AMD,” said Ruben van Dongen, head of academic solutions and business development. “Of course, we want to be friends with the entire industry. We are neutral, truly neutral.”

SCALE has been shipping for only about two years, so Spectral does not yet have a long track record. Spectral has around 30 employees and is expanding. It sells the compiler to commercial users while offering it free to academic and nonprofit groups. The software has already been tested on large systems, including the Frontier supercomputer at Oak Ridge National Laboratory.

For the users Spectral is targeting, the appeal is mostly practical. Rewriting large CUDA codebases for different hardware is time-consuming and resource-intensive. Spectral is pitching a simpler path. “Especially in the field of research, the researchers lack time,” van Dongen said. “Instead of having to rewrite the entire code base or port away from their current existing code base, they can just recompile with our solution and even increase performance benefits.”

Spectral is stepping into a market dominated by Nvidia at a time when demand for GPU and AI infrastructure is rising quickly. Spectral’s approach hinges on a simple idea: keep CUDA as the standard, but break its dependence on a single vendor’s hardware.

More than three years after ChatGPT’s launch brought generative AI into the mainstream, OpenAI is broadening its focus beyond individual users to families.

OpenAI is hiring a dedicated product manager in San Francisco to build experiences for families, caregivers, and older adults across its products. The role calls for experience building products for parents and families, and other trust-sensitive consumer experiences, according to the job posting.

The hiring comes as ChatGPT’s audience continues to broaden beyond younger users. According to Sensor Tower estimates shared exclusively with TechCrunch, the share of ChatGPT users aged 35 and older globally rose to 31% in Q2 from 26% a year earlier, while the share of users aged 18 to 24 fell to 29% from 34%. In the U.S., nearly one in four smartphone users who are parents used ChatGPT during the quarter, up from 16% a year earlier, the firm estimates.

OpenAI did not respond to requests for comment about the job posting.

A dedicated product role focused on families signals that OpenAI is beginning to think about its products less as tools for individual productivity and more as technology designed for households, said Ben Bajarin, chief executive of technology consultancy Creative Strategies.

“This is similar to the path Google, Apple, and Meta eventually followed as their platforms became embedded in everyday life, but AI raises the stakes because the assistant is not just mediating content or devices,” he told TechCrunch.

That shift also brings new trust and safety challenges. Stephen Balkam, chief executive of the Family Online Safety Institute, said the hiring reflects both the maturation of OpenAI and a growing recognition that AI products used by children and teenagers require different safeguards than those designed for adults.

“I see this as safety by redesign,” Balkam told TechCrunch. “You take the initial product or service that was released… not really with kids in mind… so this is a much-needed reaction and response.”

The comments come as new research published this week by the Family Online Safety Institute found that parents are underestimating how often their children use generative AI. While 27% of U.S. parents said their child had used generative AI in the past week, 38% of children reported doing so themselves, according to the survey of more than 4,000 families in the United States and Australia.

Balkam told TechCrunch that AI companies should build products differently for younger users, with stronger content controls, age-appropriate experiences, parental oversight, and reminders to inform users that they are interacting with an AI — and not a human.

The hiring also comes amid growing scrutiny of how AI companies protect younger users. OpenAI has faced multiple lawsuits from parents alleging that ChatGPT contributed to harm suffered by their children, including in cases involving suicide.

In response to some of those concerns, OpenAI has introduced a series of safety measures over the past year, including parental controls for teen accounts, routing sensitive conversations to reasoning models designed to better handle signs of distress, and, more recently, an optional “Trusted Contact” feature that can alert a family member or caregiver in cases of potential self-harm.

AI companies, Balkam said, have an opportunity to avoid the mistakes made by social media platforms, which for years treated children much like adults before adding stronger safeguards amid mounting public pressure and regulatory scrutiny.

The hiring also aligns with OpenAI’s broader efforts around families. In a recent workshop organized with the San Antonio Spurs Community Impact organization and the Positive Coaching Alliance, the company said it aimed to explore AI’s role in learning, coaching, and youth engagement.

That said, the demographic shift is not unique to ChatGPT, though OpenAI’s audience is changing in some distinct ways.

Sensor Tower estimates that users aged 25 to 34 account for 40% of the global app audiences for Anthropic’s Claude and Google’s Gemini, matching ChatGPT, compared with 33% for Microsoft’s Copilot. Copilot, however, skews older, with 20% of its users aged 45 and above, compared with 14% for Claude, 12% for Gemini, and 11% for ChatGPT.

While ChatGPT remains relatively underpenetrated among older users, it is adding them faster than its rivals. The share of users aged 45 and above rose three percentage points year-over-year in the second quarter, compared with a two-point increase for Copilot and declines for Claude and Gemini, according to Sensor Tower.

Among U.S. smartphone users who are parents, Gemini had the widest reach at 32% in Q2, followed by ChatGPT at 24%, Claude at 4%, and Copilot at 2%.

For Bajarin, OpenAI’s decision to hire a product manager focused on families signals where consumer AI is headed. As AI becomes a technology shared across generations, he expects companies to roll out family plans, child and teen profiles, caregiver tools, shared household memory, AI tutoring, and stronger safety controls.

Netflix executives have reportedly discussed adding always-on live channels to the service. The channels would run genre-based programming around the clock, all comedies or all action films, according to a Wall Street Journal report relayed by The Verge.

They have also discussed folding rival streaming subscriptions into Netflix itself. Peacock was named specifically, appearing as a tile on the Netflix home page and billed through Netflix.

Neither is a product. Both are internal conversations, with no launch date, pricing, or confirmation from the company.

Yes, this is cable

The comparison writes itself. A grid of always-on genre channels bundled with other people’s services, billed on one invoice, is a fairly precise description of the thing Netflix spent two decades dismantling.

The bundling half is not novel. Amazon Prime Video and Apple TV+ have sold third-party subscriptions as add-ons for years, and it is a reliable way to take a cut of someone else’s revenue.

The channels half is more interesting, because it addresses a problem Netflix built itself.

The problem with infinite choice

Netflix made browsing exhausting. The library grew so large that picking something became the hardest part of the evening, and the company has already turned to generative AI to fix the content overload it helped create.

A channel solves that instantly. It removes the decision entirely, which is precisely why linear television survived as long as it did.

It is also a defensive move. Free ad-supported services like Pluto TV and Tubi have captured exactly this casual, low-effort viewing, and Paramount is consolidating Pluto TV onto a unified stack to press the advantage.

The advertising logic underneath

Follow the inventory. A channel that never stops is a stream of unskippable ad slots, which a browse-and-pick catalogue is not.

That matters more than it used to. Netflix has been chasing attention wherever it lives, striking short-form video deals with publishers to compete with YouTube and TikTok.

The financial context is not comfortable either. The company authorised a $25bn share buyback after its stock fell 10%, which is what a business does when it wants to reassure people.

The bit worth watching

Always-on channels are, by design, engagement machinery. They exist to keep the screen on after the thing you chose has finished, and to make stopping require an act of will.

Netflix is currently defending a lawsuit from the Texas attorney general alleging addictive design and improper data collection, claims the company disputes and which remain unproven. Shipping a feature explicitly built to reduce the friction of not stopping is, at minimum, awkward timing.

None of this makes the idea bad. Removing decision fatigue is a real service, and plenty of people want the television equivalent of a radio station.

But it is worth naming the trade. Netflix disrupted cable by giving viewers control, and it may be about to discover that control was never what most viewers wanted, only what they said they wanted.

DeepSeek’s recent decision to drastically cut pricing on its V4-Pro model by 75% should have been unequivocally good news for enterprise AI vendors and developers. Instead, many are discovering that cheaper models don’t automatically translate into healthier margins.

The reason is simple: While inference costs plummet, agent systems are voraciously consuming tokens faster than prices are declining. For the last 2 decades, software economics was dictated by the same rule. Infra became cheaper every year whereas applications became more capable. AI was initially hypothesized to follow the same pattern. As frontier models improved and token prices dropped, many assumed inference would become a negligible operating expense.That assumption has begun crumbling exponentially. 

A chatbot usually turns one user question into one model call. An agent turns it into a chain of planning, retrieval, tool use, verification, summarization, and follow-up decisions. The user sees one answer. The vendor pays for the loop. That is the 100x problem: The same user-visible request can cost a lot  more to serve as an agentic workflow than as a chatbot or retrieval-augmented generation (RAG) response. In longer-running workflows, the multiplier is higher. Falling model prices help, but they do not fix a product architecture that turns one prompt into dozens of billable operations.

The scale of what is now at stake is clear in how model providers themselves are pricing developer relationships. OpenAI’s proposed program to give every Y Combinator startup $2 million in API credits — a number that would have funded an entire seed round in any prior tech cycle, and when the same cohort got by on a few thousand dollars of AWS credits — is less a recruiting perk than an admission of what it now costs to run an AI-native company through its first year of product. For established enterprises retrofitting agents into existing product lines, the absolute numbers are larger still.

What token amplification is

In a single-turn chatbot, one user message produces roughly one model call. Input-to-billed ratio is about 1:5.

In a multi-step agent rolled out across customer support, sales operations, finance, legal review, and engineering, that ratio routinely lands at 1:700 or higher. Every loop iteration carries forward the cumulative conversation, tool outputs, and reasoning traces. Each step appends; nothing is dropped.

A “simple” agent query like “What did our top customer ask about last week?” typically touches seven priced operations before returning an answer:

1. User prompt (~50 tokens)

2. System prompt and tool definitions (~3,000 tokens, repeated on every call)

3. Retrieval (~5,000 tokens of context)

4. Model call #1 — tool selection (8,000 in / 200 out)

5. Tool execution (~4,000 tokens returned)

6. Model call #2 — summarization (12,000 in / 400 out)

7. Model call #3 — follow-up decision (12,400 in / 100 out)

One sentence in, roughly 35,000 input tokens billed. Somewhere between $0.10 and $0.40 per query on a frontier model. Multiply that by a million queries a month — the table-stakes volume for any enterprise B2B feature — and the line item is six figures.

Why this breaks the existing AI business model

The dominant pricing story for enterprise AI has been seat-based SaaS: Pay per-user per-month, deliver agent capability, capture margin. That model assumes a reasonably bounded cost-per-user.

Token amplification breaks the assumption. A power user running 50 agent invocations a day on a $40/seat plan can cost more in inference than the plan charges. Token amplification shatters the traditional SaaS pricing model. When a power user’s daily agent activity costs more in inference than their monthly subscription fee, vendor gross margins turn negative, a paradox that compounds as customers deepen their agent adoption, the very usage curve vendors are selling to their boards. Several vendors are now privately reporting negative gross margins on heavy users, mirroring recent cloud expenditure reports from the Bessemer ‘Supernova’ cohort, where the correlation between AI-agent adoption and gross margin contraction has moved from a theoretical risk to a primary P&L headwind.

The visible symptoms have started leaking into public coverage. Bloomberg this week documented a widening gap between Salesforce’s Agentforce marketing demos and the capabilities actually shipping to customers. This is the kind of gap that opens predictably when promised functionality is technically possible but uneconomical to serve at the price the seat plan implies. Salesforce is the most-watched case, not a unique one.

“For my team, the cost of compute is far beyond the costs of the employees.” — Bryan Catanzaro, VP of Applied Deep Learning, Nvidia

The strategic implication is not “AI is expensive.” It is that the dominant business model assumed by most AI-native company plans does not survive contact with agentic workloads.

A simple example

Consider an enterprise software vendor charging $40 per-user per-month for an AI-enabled support assistant. A traditional chatbot might cost only a few cents per user per day in inference, leaving healthy gross margins.

Now replace that chatbot with a fully agentic workflow capable of investigating tickets, querying internal systems, drafting responses, validating outputs, and escalating exceptions. If a heavy user executes 50 to 100 agent requests per day, inference consumption can increase by an order of magnitude. What was once a negligible infrastructure cost becomes a material operating expense.

This creates an unusual dynamic: The customers receiving the most value from the product are often the customers generating the highest inference costs. In extreme cases, vendors can find themselves with their most engaged users contributing the least profit. The result is a growing realization across enterprise software that agent adoption and margin expansion are no longer automatically aligned.

Agent orchestration is the new moat

The technical responses are known and converging. They are not novel, but they are critical for survival

• Cost-aware routing: This technique involves a small classifier model that decides which tier (Haiku, Sonnet, Opus equivalents) handles each query. Well-tuned routers cut inference bills by around 60% without any degradation in quality

• Prompt caching: Anthropic, OpenAI, and Google now offer 75 to 90% discounts on cached prefixes. 

• Context discipline: You can truncate tool outputs, prune reasoning traces, and cap tool depth to prevent your agent from going down a rabbit hole

• Speculative decoding: for self-hosted deployments, this technique guarantees 2 to 3X effective throughput on the same GPUs.

“Organizations using orchestration-led governance report stronger productivity gains — a holistic orchestration layer is associated with six times greater productivity impact than compliance‑only approaches” — IBM

The companies building this layer well are starting to look less like microservice operators and more like financial trading systems: Every routing decision priced, every path with its own P&L, every tenant on a metered budget.

What enterprise leaders should actually do

Four moves separate the companies that will still have margin in 24 months from the ones that won’t:

1. Make inference cost a first-class metric. Track it per-feature, per-tenant, per-query class the same way cloud cost was tracked starting in the mid-2010s.

2. Budget like a media buyer. Set cost-per-thousand-queries ceilings per feature. Cap them. Alert on overruns. Engineering will not enforce this on its own.

3. Treat the router as core infrastructure, not an optimization. It is the new load balancer.

4. Audit prompts quarterly. A 4,000-token system prompt that grew organically over six months is a six-figure bill in slow motion. Most teams have never read their own production prompts end to end.

5. Negotiate volume commits early. Frontier-model vendors now offer reserved-instance-style prepaid commits at substantial discounts. List price is the worst price any enterprise will ever pay.

The next 24 months

The structural shift underneath agentic AI is not that it is expensive. As DeepSeek’s price cut today underscores, frontier inference unit costs are dropping roughly 3X per year, and the curve is not slowing.

The shift is that amplification is outrunning the price cuts. Cutting per-token costs 75% does not help a company whose agents are doing 700X more tokens per user query than its pricing model assumed. For the first time since the cloud era began, architecture decisions are again financial decisions in real time. A prompt redesign is a margin event. A poorly bound agent loop is an outage with a credit card attached.

The companies that survive the next 24 months of AI infrastructure pricing will not be the ones running the cheapest model. They will be the ones whose agents are smart and know what they cost to think.

That is the 100X problem. And it is arriving faster than the price cuts can hide it.

Maitreyi Chatterjee is a senior software engineer at a big tech company.

Devansh Agarwal works as an ML engineer at a leading tech company.

Welcome to the VentureBeat community!

Our guest posting program is where technical experts share insights and provide neutral, non-vested deep dives on AI, data infrastructure, cybersecurity and other cutting-edge technologies shaping the future of enterprise.

Read more from our guest post program — and check out our guidelines if you’re interested in contributing an article of your own!

Off a CRT television, the footage captures Mario mid-jump in Bob-omb Battlefield. The camera follows him across familiar terrain. No emulator sits between the console and the screen. This moment marks real hardware execution after roughly 3000 separate builds and months of relentless iteration by one dedicated developer. The project began as a fork of earlier work by malucard, who first adapted the open-source Super Mario 64 decompilation for PlayStation 1 targets.

Early versions showed some promise, but they were rough around the edges, with trees floating around, animations stalling or breaking, and the camera refusing to act properly. In several instances, performance was sluggish. Still, these early builds demonstrated that the concept could be made to work, even if it wasn’t remotely playable on a console; however, that’s where Eyepatch Entertainment stepped in, and they decided to take on the challenge of getting it stable and running smoothly on a PlayStation’s far-from-generous 2MB of RAM.

Nintendo Switch 2: Choose Your Game Bundle

• Kick off the fun with a Nintendo Switch 2 system, your choice of a select digital game, and a savings* of up to $29.99!
• Includes choice of either the Mario Kart World, Donkey Kong Bananza, or Pokémon Pokopia digital game download
• One system, three play modes: TV, Tabletop, and Handheld

Super Mario 64, of course, was initially intended for the Nintendo 64, which had more memory and rasterization processors. In contrast, the PlayStation ran a different CPU, had a simpler graphics system, and used a disc rather than a cartridge. Every image, procedure, and byte of data had to be rethought to function inside the PlayStation’s limited memory space to prevent the game from crashing all the time.

The first round of work was focused on getting the game to boot properly and not run out of memory. That’s a significant issue, and levels like Bob-omb Battlefield were practically impossible to work on owing to allocation concerns. The game occasionally crashed, and when it did, it was limited to a few frames per second. You have to fix the CD image handling to remedy the sector alignment and controller mode issues that were causing texture corruption. It was difficult to debug at first, simply comparing screenshots to what we expected to happen, but after he got GDB and backtracing using the PCSX-Redux emulator, things improved significantly.

Then, in a big breakthrough, he successfully got the game to run on an actual PlayStation. Now, I understand that sounds simple, but getting it to do anything close to what we saw in the emulator was a huge feat. Of course, it wasn’t all smooth sailing; a few additional concerns surfaced. For example, because of the way the framebuffer was flipped, Mario’s model was allowing in light from behind, causing the game to drop frames and freeze. So he fixed that and a few other bugs, and the game began to take shape. He changed the topography to just build the necessary level detail and simplified elements of the 3D models, such as Mario’s legs. This made a big impact in memory and performance.

Audio, oh boy, was a whole other ballgame. On an actual PlayStation, he would occasionally hear stillness or a lot of crackling. The music was encoded as 8-bit ADPCM, which worked flawlessly on the emulator, however the real device required 4-bit XA-ADPCM. We had to re-encode the entire audio. We changed the sector interleaving so that the PlayStation could read it properly, and we even included additional automated checks to ensure that the same issue did not occur again.

The good news is that the most recent builds are beginning to resemble what he envisioned. The skybox, with its flowing clouds and ocean backdrop, is operating perfectly. The Peach sequence is playing out exactly how he had imagined. After very rigorous tuning, Bob-omb Battlefield now runs at a consistent 20 to 28 frames per second. The Chain Chomp is loading properly, animating correctly, and even demolishing its gate on time. Even the door and painting transitions function without crashing the game, and the castle interior runs nicely once we set the frame rate to 30 FPS to avoid timing difficulties.