Apple execs are now acknowledging the company is going to need to raises prices to cope with higher component costs, particularly memory. Apple has been battling to avoid doing so, but warned this week that those efforts are “unsustainable.”

The price pressure is being felt across the tech industry. Omdia predicts the average selling price (ASP) of smartphones globally will increase by around 20% this year — partly because AI industry demands have pushed up RAM prices and partly because of rapid cost increases in processor production as a direct consequence of war in the Middle East. This impacts anything that uses memory or processors.

Apple must raise prices, CEO warns

“We’re doing our best to mitigate the huge increases that are being passed to us, and we’ve been trying to shield our customers from the increases, but the situation has become unsustainable,” current Apple CEO Tim Cook told the Wall Street Journal.

OK, so prices are going to go up. But how?

If Apple must raise prices, the way it does so matters. Will it increase prices across the board, or choose a subtle, strategic approach that protects margins while also maintaining rapidly accelerating market share growth?

Until now, Apple has been benefiting from the rising tide in memory prices. While competitors have been steadily increasing prices across the board, Apple introduced its first $599 laptop even as it eats into mid-range smartphone sales. 

To grow, or to grow less?

The PC industry is not growing (outside of Cupertino), and Apple and its investors already know that the inordinately high customer satisfaction ratings its platforms enjoy mean that when people switch to them they tend to stick around. That’s why it matters that Cook highlighted record switchers and first-time platform users during the company’s most recent fiscal call. Apple is growing new customers, people who are likely to sign-up for services, accessories, and future upgrades. Winning the customer is an investment in future performance.

When it comes to increasing prices, we’ve seen some gentle experimentation in the Mac mini line, where the entry-level $599 model was effectively replaced by a more expensive $799 SKU. There have been suggestions Apple could do the same with the $599 MacBook Neo, dumping the entry-level build in favor of the (better) $699 model. I’m not certain that’s the best approach. 

A matter of choice

I’d argue that the $599 price point is psychologically the best bit of marketing Apple has ever done as it draws new users to look at its platforms. I also think substantial numbers of people already shift up to the $699 version, if they can, making that entry-level price an incredibly effective marketing tool with which to woo customers. It looks like that approach is working; the MacBook Neo has led the Amazon Laptop Sales chart ever since it was introduced and is selling in huge quantities globally.

Francisco Jeronimo, IDC vice president for data and analytics, sees this, writing: “It will be interesting to see whether Apple raises prices across its entire portfolio or plays strategically by subsidizing its more affordable products to gain market share and grow its installed base.”

If Apple raises prices at the price sensitive, entry-level end of its market, it runs the risk of slowing its own resurgence. Given that a 10% price hike here is only worth around $60, the revenue generated might not justify blunting Apple’s offer to new users.

Pricing and strategy

But Apple isn’t defined by its entry-level markets. Its biggest earning opportunities continue to be generated by the aspirational middle class, who, while feeling the pinch of declining living standards in many markets, can still handle the price of a top-of-the-range iPhone Pro. A 10% price increase at that end of the market will generate a much more significant chunk of revenue – and since rival vendors are raising their prices at this end of the market, Apple can follow suit while still seeming affordable. It’s an approach that maintains growth while optimizing margins.

Apple generates roughly half its revenue from iPhone sales, with the Pro and Pro Max devices accounting for a significant chunk of that money. Consumers looking to buy one of Apple’s high-end devices are likely to be less price-sensitive than those making purchases at the low- and mid-points of the company’s market range. 

With that in mind, it makes more sense for the company to raise prices at the top, while working to maintain the cost of entry at the lower end. It could, of course, raise prices universally, but doing so makes less sense.

What next?

Whatever strategy Apple adopts, company leadership is clear that component pricing is hitting hard. “We definitely need memory pricing and supply to return to reasonable levels for consumer products. That’s the bottom line,” Cook said.

You can follow me on social media! Join me on BlueSky,  LinkedIn, Mastodon and subscribe to The Core for your daily fix of Apple News.

France’s OVHcloud is moving beyond cloud infrastructure into frontier AI model development, a shift that could test whether Europe can produce another serious alternative to US and Chinese AI systems.

The company, one of Europe’s leading homegrown cloud providers, plans to train a family of models from scratch and aims to open-source them once they meet its performance targets, CEO Octave Klaba told Reuters.

The move would put OVHcloud in closer comparison with Mistral AI, the Paris-based model developer that has become Europe’s most visible challenger to US AI labs.

Klaba said the economics of building advanced AI models have changed, with improvements in chips, training methods, and synthetic data reducing the cost of a project that may once have required about $1.15 billion (€1 billion) to now cost less than $230 million (€200 million).

Reuters reported that OVHcloud said one of its models has completed pre-training on Jupiter, the Germany-based EuroHPC supercomputer described as Europe’s fastest and its first exascale system, though the company has not yet disclosed detailed performance benchmarks.

This comes as European governments and enterprises are increasingly having to assess AI infrastructure through the lens of data governance and continuity of access, rather than performance alone.

Those concerns were sharpened this month after Anthropic said a US government export-control directive required it to suspend access to its Fable 5 and Mythos 5 models by foreign nationals inside and outside the US.

Training is only the opening cost

OVHcloud’s lower cost estimate does not capture the full cost of becoming a frontier AI model provider, said Neil Shah, vice president for research and partner at Counterpoint Research.

The $230 million (€200 million) figure likely refers mainly to the initial training run, Shah said. Once trained, however, models require continued investment because they can become depreciating assets if they are not improved with fresh data.

OVHcloud would also need to spend on fine-tuning, post-training, sovereign infrastructure, storage, security, distribution, and enterprise support. It would also need enough scale to make model serving economically viable against established AI providers such as Google and Anthropic.

“Model is seen as a depreciating asset if it is not consistently trained and kept fresh with the data,” Shah said.

That makes OVHcloud’s plan a test not only of technical capability, but also of policy support and economic viability. If the company falls short, enterprises may be reluctant to shift workloads away from more established models.

The lower training cost could still give OVHcloud a credible starting point, said Charlie Dai, principal analyst at Forrester.

The budget range can be enough to produce a credible frontier model as efficiency gains reduce the cost of entry, Dai said. But enterprise competitiveness will depend on sustained capabilities beyond training, including inference efficiency, data pipelines, evaluation frameworks, and ecosystem reach.

Buyers need proof

OVHcloud’s plan remains an expression of intent rather than demonstrated capability, said Sanchit Vir Gogia, chief analyst at Greyhound Research, pointing to the absence of published benchmarks and other details.

“$200 million now buys a serious training run,” Gogia said. “It does not buy a serious enterprise AI franchise.”

Gogia said questions around sovereignty also extend to the infrastructure used to train the model, noting that pre-training was run on Jupiter rather than on infrastructure owned or controlled by OVHcloud.

The system is a publicly owned European supercomputer in Germany that runs on American silicon, Gogia said, adding that this shows how partial European AI sovereignty remains.

CIOs will need evidence that the models can be supported in production, governed effectively, audited when needed, and exited without major disruption.

Gogia said a European-owned model could reduce some dependence on US and Chinese providers, but would not remove jurisdictional risk. “Sovereignty does not abolish the off switch,” he said. “It changes whose hand rests upon it.”

OVHcloud’s move into model development could also alter the lock-in risks enterprises need to assess, Gogia said. Customers may be able to move cloud infrastructure later, but find it harder to shift AI workloads once applications and processes are built around a provider’s models and governance tools.

What a long, strange trip it’s been.

From its inaugural release to today, Android has transformed visually, conceptually and functionally — time and time again. Google’s mobile operating system may have started out scrappy, but holy moly, has it ever evolved.

Here’s a fast-paced tour of Android version highlights from the platform’s birth to present. (Feel free to skip ahead if you just want to see what’s new in the most recent Android 17 update.)

Android versions 1.0 to 1.1: The early days

Android made its official public debut in 2008 with Android 1.0 — a release so ancient it didn’t even have a cute codename.

Things were pretty basic back then, but the software did include a suite of early Google apps like Gmail, Maps, Calendar, and YouTube, all of which were integrated into the operating system — a stark contrast to the more easily updatable standalone-app model employed today.

Android version 1.5: Cupcake

With early 2009’s Android 1.5 Cupcake release, the tradition of Android version names was born. Cupcake introduced numerous refinements to the Android interface, including the first on-screen keyboard — something that’d be necessary as phones moved away from the once-ubiquitous physical keyboard model.

Cupcake also brought about the framework for third-party app widgets, which would quickly turn into one of Android’s most distinguishing elements, and it provided the platform’s first-ever option for video recording.

Android version 1.6: Donut

Android 1.6, Donut, rolled into the world in the fall of 2009. Donut filled in some important holes in Android’s center, including the ability for the OS to operate on a variety of different screen sizes and resolutions — a factor that’d be critical in the years to come. It also added support for CDMA networks like Verizon, which would play a key role in Android’s imminent explosion.

Android versions 2.0 to 2.1: Eclair

Keeping up the breakneck release pace of Android’s early years, Android 2.0, Eclair, emerged just six weeks after Donut; its “point-one” update, also called Eclair, came out a couple months later. Eclair was the first Android release to enter mainstream consciousness thanks to the original Motorola Droid phone and the massive Verizon-led marketing campaign surrounding it.

Verizon’s “iDon’t” ad for the Droid.

The release’s most transformative element was the addition of voice-guided turn-by-turn navigation and real-time traffic info — something previously unheard of (and still essentially unmatched) in the smartphone world. Navigation aside, Eclair brought live wallpapers to Android as well as the platform’s first speech-to-text function. And it made waves for injecting the once-iOS-exclusive pinch-to-zoom capability into Android — a move often seen as the spark that ignited Apple’s long-lasting “thermonuclear war” against Google.

Android version 2.2: Froyo

Just four months after Android 2.1 arrived, Google served up Android 2.2, Froyo, which revolved largely around under-the-hood performance improvements.

Froyo did deliver some important front-facing features, though, including the addition of the now-standard dock at the bottom of the home screen as well as the first incarnation of Voice Actions, which allowed you to perform basic functions like getting directions and making notes by tapping an icon and then speaking a command.

Notably, Froyo also brought support for Flash to Android’s web browser — an option that was significant both because of the widespread use of Flash at the time and because of Apple’s adamant stance against supporting it on its own mobile devices. Apple would eventually win, of course, and Flash would become far less common. But back when it was still everywhere, being able to access the full web without any black holes was a genuine advantage only Android could offer.

Android version 2.3: Gingerbread

Android’s first true visual identity started coming into focus with 2010’s Gingerbread release. Bright green had long been the color of Android’s robot mascot, and with Gingerbread, it became an integral part of the operating system’s appearance. Black and green seeped all over the UI as Android started its slow march toward distinctive design.

Android 3.0 to 3.2: Honeycomb

2011’s Honeycomb period was a weird time for Android. Android 3.0 came into the world as a tablet-only release to accompany the launch of the Motorola Xoom, and through the subsequent 3.1 and 3.2 updates, it remained a tablet-exclusive (and closed-source) entity.

Under the guidance of newly arrived design chief Matias Duarte, Honeycomb introduced a dramatically reimagined UI for Android. It had a space-like “holographic” design that traded the platform’s trademark green for blue and placed an emphasis on making the most of a tablet’s screen space.

While the concept of a tablet-specific interface didn’t last long, many of Honeycomb’s ideas laid the groundwork for the Android we know today. The software was the first to use on-screen buttons for Android’s main navigational commands; it marked the beginning of the end for the permanent overflow-menu button; and it introduced the concept of a card-like UI with its take on the Recent Apps list.

Android version 4.0: Ice Cream Sandwich

With Honeycomb acting as the bridge from old to new, Ice Cream Sandwich — also released in 2011 — served as the platform’s official entry into the era of modern design. The release refined the visual concepts introduced with Honeycomb and reunited tablets and phones with a single, unified UI vision.

ICS dropped much of Honeycomb’s “holographic” appearance but kept its use of blue as a system-wide highlight. And it carried over core system elements like on-screen buttons and a card-like appearance for app-switching.

Android 4.0 also made swiping a more integral method of getting around the operating system, with the then-revolutionary-feeling ability to swipe away things like notifications and recent apps. And it started the slow process of bringing a standardized design framework — known as “Holo” — all throughout the OS and into Android’s app ecosystem.

Android versions 4.1 to 4.3: Jelly Bean

Spread across three impactful Android versions, 2012 and 2013’s Jelly Bean releases took ICS’s fresh foundation and made meaningful strides in fine-tuning and building upon it. The releases added plenty of poise and polish into the operating system and went a long way in making Android more inviting for the average user.

Visuals aside, Jelly Bean brought about our first taste of Google Now — the spectacular predictive-intelligence utility that’s sadly since devolved into a glorified news feed. It gave us expandable and interactive notifications, an expanded voice search system, and a more advanced system for displaying search results in general, with a focus on card-based results that attempted to answer questions directly.

Multiuser support also came into play, albeit on tablets only at this point, and an early version of Android’s Quick Settings panel made its first appearance. Jelly Bean ushered in a heavily hyped system for placing widgets on your lock screen, too — one that, like so many Android features over the years, quietly disappeared a couple years later.

Android version 4.4: KitKat

Late-2013’s KitKat release marked the end of Android’s dark era, as the blacks of Gingerbread and the blues of Honeycomb finally made their way out of the operating system. Lighter backgrounds and more neutral highlights took their places, with a transparent status bar and white icons giving the OS a more contemporary appearance.

Android 4.4 also saw the first version of “OK, Google” support — but in KitKat, the hands-free activation prompt worked only when your screen was already on and you were either at your home screen or inside the Google app.

The release was Google’s first foray into claiming a full panel of the home screen for its services, too — at least, for users of its own Nexus phones and those who chose to download its first-ever standalone launcher.

Android versions 5.0 and 5.1: Lollipop

Google essentially reinvented Android — again — with its Android 5.0 Lollipop release in the fall of 2014. Lollipop launched the still-present-today Material Design standard, which brought a whole new look that extended across all of Android, its apps and even other Google products.

The card-based concept that had been scattered throughout Android became a core UI pattern — one that would guide the appearance of everything from notifications, which now showed up on the lock screen for at-a-glance access, to the Recent Apps list, which took on an unabashedly card-based appearance.

Lollipop introduced a slew of new features into Android, including truly hands-free voice control via the “OK, Google” command, support for multiple users on phones and a priority mode for better notification management. It changed so much, unfortunately, that it also introduced a bunch of troubling bugs, many of which wouldn’t be fully ironed out until the following year’s 5.1 release.

Android version 6.0: Marshmallow

In the grand scheme of things, 2015’s Marshmallow was a fairly minor Android release — one that seemed more like a 0.1-level update than anything deserving of a full number bump. But it started the trend of Google releasing one major Android version per year and that version always receiving its own whole number.

Marshmallow’s most attention-grabbing element was a screen-search feature called Now On Tap — something that, as I said at the time, had tons of potential that wasn’t fully tapped. Google never quite perfected the system and ended up quietly retiring its brand and moving it out of the forefront the following year.

Android 6.0 did introduce some stuff with lasting impact, though, including more granular app permissions, support for fingerprint readers, and support for USB-C.

Android versions 7.0 and 7.1: Nougat

Google’s 2016 Android Nougat releases provided Android with a native split-screen mode, a new bundled-by-app system for organizing notifications, and a Data Saver feature. Nougat added some smaller but still significant features, too, like an Alt-Tab-like shortcut for snapping between apps.

Perhaps most pivotal among Nougat’s enhancements, however, was the launch of the Google Assistant — which came alongside the announcement of Google’s first fully self-made phone, the Pixel, about two months after Nougat’s debut. The Assistant would go on to become a critical component of Android and most other Google products and is arguably the company’s foremost effort today.

Android version 8.0 and 8.1: Oreo

Android Oreo added a variety of niceties to the platform, including a native picture-in-picture mode, a notification snoozing option, and notification channels that offer fine control over how apps can alert you.

The 2017 release also included some noteworthy elements that furthered Google’s goal of aligning Android and Chrome OS and improving the experience of using Android apps on Chromebooks, and it was the first Android version to feature Project Treble — an ambitious effort to create a modular base for Android’s code with the hope of making it easier for device-makers to provide timely software updates.

Android version 9: Pie

The freshly baked scent of Android Pie, a.k.a. Android 9, wafted into the Android ecosystem in August of 2018. Pie’s most transformative change was its hybrid gesture/button navigation system, which traded Android’s traditional Back, Home, and Overview keys for a large, multifunctional Home button and a small Back button that appeared alongside it as needed.

Pie included some noteworthy productivity features, too, such as a universal suggested-reply system for messaging notifications, a new dashboard of Digital Wellbeing controls, and more intelligent systems for power and screen brightness management. And, of course, there was no shortage of smaller but still-significant advancements hidden throughout Pie’s filling, including a smarter way to handle Wi-Fi hotspots, a welcome twist to Android’s Battery Saver mode, and a variety of privacy and security enhancements.

Android version 10

Google released Android 10 — the first Android version to shed its letter and be known simply by a number, with no dessert-themed moniker attached — in September of 2019. Most noticeably, the software brought about a totally reimagined interface for Android gestures, this time doing away with the tappable Back button altogether and relying on a completely swipe-driven approach to system navigation.

Android 10 packed plenty of other quietly important improvements, including an updated permissions system with more granular control over location data along with a new system-wide dark theme, a new distraction-limiting Focus Mode, and a new on-demand live captioning system for any actively playing media.

Android version 11

Android 11, launched at the start of September 2020, was a pretty substantial Android update both under the hood and on the surface. The version’s most significant changes revolve around privacy: The update built upon the expanded permissions system introduced in Android 10 and added in the option to grant apps location, camera, and microphone permissions only on a limited, single-use basis.

Android 11 also made it more difficult for apps to request the ability to detect your location in the background, and it introduced a feature that automatically revokes permissions from any apps you haven’t opened lately. On the interface level, Android 11 included a refined approach to conversation-related notifications along with a new streamlined media player, a new Notification History section, a native screen-recording feature, and a system-level menu of connected-device controls.

Android version 12

Google officially launched the final version of Android 12 in October 2021, alongside the launch of its Pixel 6 and Pixel 6 Pro phones.

In a twist from the previous several Android versions, the most significant progressions with Android 12 were mostly on the surface. Android 12 featured the biggest reimagining of Android’s interface since 2014’s Android 5.0 (Lollipop) version, with an updated design standard known as Material You — which revolves around the idea of you customizing the appearance of your device with dynamically generated themes based on your current wallpaper colors. Those themes automatically change anytime your wallpaper changes, and they extend throughout the entire operating system interface and even into the interfaces of apps that support the standard.

Surface-level elements aside, Android 12 brought a (long overdue) renewed focus to Android’s widget system along with a host of important foundational enhancements in the areas of performance, security, and privacy. The update provided more powerful and accessible controls over how different apps are using your data and how much information you allow apps to access, for instance, and it included a new isolated section of the operating system that allows AI features to operate entirely on a device, without any potential for network access or data exposure.

Android version 13

Android 13, launched in August 2022, was simultaneously one of the most ambitious updates in Android history and one of the most subtle version changes to date.

On tablets and foldable phones, Android 13 introduced a slew of significant interface updates and additions aimed at improving the large-screen Android experience — including an enhanced split-screen mode for multitasking and a ChromeOS-like taskbar for easy app access from anywhere.

On regular phones, Android 13 brought about far less noticeable changes — mostly just some enhancements to the system clipboard interface, a new native QR code scanning function within the Android Quick Settings area, and a smattering of under-the-hood improvements.

Android version 14

Following a full eight months of out-in-the-open refinement, Google’s 14th Android version landed at the start of October 2023, in the midst of the company’s Pixel 8 and Pixel 8 Pro launch event.

Like the version before it, Android 14 didn’t look like much on the surface. That’s in part because of the trend of Google moving more and more toward a development cycle that revolves around smaller ongoing updates to individual system-level elements year-round — something that’s actually a significant advantage for Android users, even if it does have an awkward effect on people’s perception of progress.

But despite the subtle nature of its first impression, Android 14 delivered a fair amount of noteworthy new goodies. The software introduced a new system for dragging and dropping text between apps, for instance, as well as a number of new improvements to privacy and security — including a new settings-integrated dashboard for managing health and fitness data and a more info-rich and context-requiring system for seeing exactly why apps want access to your location. And it brought about a new set of native customization options for the Android lock screen.

Android version 15

Though Android 15 followed the trend of significant advancements arriving as their own separate rollouts — outside of and even ahead of its arrival, as an official operating system update — 2024’s new Android version was certainly no slouch.

The software introduced a number of noteworthy new features — including a redesigned system volume panel, an option to automatically re-enable a device’s Bluetooth radio a day after it’s been disabled, and a Pixel-specific Adaptive Vibration feature that intelligently adjusts a phone’s vibration intensity based on the environment. It also marked the debut of a system-level Private Space area that lets you keep sensitive apps out of sight and accessible only with authentication.

Add in handy touches like a space-saving app archiving option and a predictive back visual that lets you sneak a peek at where you’re headed before you get there, and this small-seeming update shaped up to be a pretty hefty progression.

Android version 16

In a marked change from recent Android upgrade cycles, Google decided to go with two new Android versions per year as of 2025 — starting with Android 16 in the spring and then following that with a smaller release in the fall.

True to that promise, Android 16 catapulted into the world in early June, creating the framework for future-facing systems such as Live Updates — a new type of notification designed to support persistent, ongoing alerts, similar to what Apple does with iOS’s Live Activities — and introducing an Advanced Protection security supermode that provides a simple single-switch way to activate a whole slew of advisable Android security settings in one fell swoop.

The update included a sprawling series of other new security strengtheners, too, making protection seem like the true centerpiece of Android 16 — even if other touches, such as a more advanced standard for hearing aid support, helped flesh out the software into a rounded and feature-rich release.

Android version 17

With its relatively low-key arrival in June 2026, Android 17 officially brings the long under-development Bubbles multitasking system to the Android-owning masses — adding an interesting new way to keep any app available on demand in a floating, collapsible window for easy ongoing access.

Speaking of bubbliness, Android 17 also includes the creator-aimed option of showing a cutout of your face from a front-facing camera over an active screen recording — because why not, right? — along with such practical touches as a more dynamic and consistent system-wide dark mode and a more nuanced and effective way to track and control app location access.

While those features and the inevitable slew of under-the-hood security, performance, and privacy improvements add up to form a compelling final picture, it’s hard not to notice that much of Google’s focus in this era is now on the AI layers surrounding Android as opposed to being on Android itself, as an operating system. The company’s I/O conference in May showcased many such measures, appropriately noting that Android was transitioning from being “an operating system” into being “an intelligence system” (whatever that means).

Most of those “intelligence system” items remain limited in ability or not yet available as of the time of Android 17’s release — like the new and improved speech-to-text system for Gboard, the custom-widget-creating system for Android phones, and the multistep automation system for allowing AI to complete complex tasks on your behalf (assuming that you (a) trust such a system to act on your behalf and (b) don’t find the level of access and resulting manner of assumptions it makes about your life to be overly creepy).

But even at its foundational level and without any AI-laden Halo effect included, Android 17 manages to hold its own — with Bubbles acting as an anchor and bringing some much-appreciated new productivity potential our way.

This article was originally published in November 2017 and most recently updated in June 2026.

Because generative AI (genAI) tools and services have become so ubiquitous (and popular), the costs of using them are going through the roof — leading to an insatiable appetite for tokens.

Tokens represent a common way to measure and price AI use. Much like letters and words in English, large language models (LLMs) grasp a sentence or query by breaking words into tokens.

With the AI explosion well under way, tokens are now “the fundamental units of data our models process, many representing a problem being solved,” according to Google CEO Sundar Pichai. (Google, by the way, processes about 3.2 quadrillion tokens a month.)

But as the price of all those tokens adds up, business and IT execs are looking for ways to cut costs while keeping corporate productivity up. Uncontrolled token use has already landed one company with an unexpected $500 million AI bill. 

There are a number of ways companies can rein in the price of AI at the model, infrastructure, silicon, and business levels. Here’s a look at how some of those savings might actually be achieved.

Switch to lower-cost models

One way of potentially saving money is by re-routing AI work to a cheaper model, Pichai said. At Google that would Gemini 3.5 Flash. It delivers “frontier-level capabilities at less than half the price of comparable frontier models.

“If companies use a mix of [Gemini 3.5] Flash and other frontier models, they could save a lot of money,” Pichai said.

Those kinds of models provide cheaper tokens, with reasoning that’s good enough for many users — if not as strong as mainstream Gemini 3.5 — to deliver useful results.

“There is sometimes overkill with the [LLMs],” said Deepak Seth, senior director analyst at Gartner. “I don’t always need a large language model which has been trained on the works of Charles Dickens and Shakespeare and Harry Potter.”

Hyperframe Research principal analyst Steven Dickens can’t stop using Amazon’s Quick, which costs $20 a month, for personal tasks. “It is great personal ROI as it has not only made tasks faster, but unlocked tasks I would never have even attempted previously,” Dickens said.

Don’t forget the hardware and software part of the equation

The token crisis isn’t new, said Dheeraj Pandey, CEO of DevRev, who likens what’s going on now in the AI market to the disruptions that emerged with the arrival of cloud computing and virtualization years ago. 

“We let chaos reign and then we had to rein in the chaos,” Pandey said. “The word that people started using was server consolidation and virtualization.”

The answer to the token problem, he said, is the same: “Anything in systems can be solved with caching and indirection.”

DevRev, for example, is building a memory layer between AI agents and primary data sources, such as Salesforce or ERP records; that can cut token load and make data movement more efficient. The layer holds a knowledge graph with answers to common agent questions and runs on cheaper CPUs, avoiding more costly GPU cycles.

Sending agents straight at systems like ServiceNow and Salesforce “will burn a lot more tokens. It’s also not precise. And finally, it’s not safe enough where I can roll it back in case an agent has committed a mistake,” Pandey said.

Network automation firm NetBrain uses a different method: It uses conventional computing to map a network’s layout then feeds only key information to models for planning and reasoning, where AI excels. “So you don’t have to spend all the tokens,” said Netbrain CTO Song Pang.

Focus on prompt efficiency

Staffing firm ManpowerGroup has found that prompt efficiency can be an effective tool for improving token use, both internally and externally for clients.

For example, users accessing its internal labor-market tool initially needed 10 follow-up questions to drill into a query. A year later, more efficient use of prompts has brought that number down to an average of four, said Max Leaming, head of data science and AI solutions at ManpowerGroup. 

“They’re using fewer tokens and they’re simply more efficient,” he said. “And that in large part has to do with your ability to prompt efficiently.”

Go local

New AI hardware that generates free tokens at home could ease some of the cost crisis.

At GTC Taipei earlier this month, Nvidia and Microsoft unveiled RTX Spark, an agentic AI desktop PC that runs agents and 120-billion-parameter models locally on Windows. The goal is “to deliver unmetered intelligence to every home and every desk with Windows,” Microsoft CEO Satya Nadella said in a statement.

Some companies are looking to reduce cloud AI costs by putting their own hardware in data centers, with vendors such as HPE and Dell providing servers installed in independent facilities. (On-premise AI is gaining ground amid sovereign AI and geopolitical concerns, including the recent conflict in the Middle East, where large data centers were struck with missiles.)

“There are local, region-specific and multiple vendor AI solutions. All of those things can help mitigate the risk. But they’re not going to eliminate it,” said Max Goss, senior director analyst at Gartner.

Use forward-deployed engineers

Reducing token costs is something that may fall to forward-deployed engineers (FDEs) in customer environments, said Taimur Rashid, managing director of AWS’s Generative AI Innovation Center.

“I expect these teams to be able to architect systems that have those cost requirements in mind, whether it’s use a different model or a different use case that doesn’t increase the per-token cost,” Rashid said.

Companies may spend heavily on token consumption, “but if you’re generating revenue, as long as the economics work out, then you’re at peace,” Rashid said.

The use of FDEs is gaining ground as IT decision-makers look to both rollout successful AI deployments while also keeping an eye on costs.

Change the measure of success from tokens to outcomes

Even with the current emphasis on reducing token use to save money, the metrics used to measure AI success are likely to shift, Gartner’s Seth said. At some point, token-based pricing will move more toward an outcome-based model, where the unit of value is outcomes, not fragments of words.

“Some companies are moving towards outcome-based pricing,” Seth said. “When people start realizing the real cost of tokens, then companies will start looking at token efficiency.”

A federal judge has rebuffed Workday’s claim that it cannot be held liable under California anti-discrimination laws when its tools are used to screen (and potentially reject) job candidates in other states.

This week, US District Judge Rita Lin indicated that she will likely allow additional state discrimination claims against Workday to move forward. This would significantly expand the closely-watched case and likely ratchet up scrutiny of AI recruiting tools and their potentially inherent biases when it comes to age, race, sex, disabilities, and other factors.

Further, it could indicate that, even if a company is not the final employer, it may be held liable if its tools materially influence who gets rejected. This could set new legal standards for AI hiring systems, and have implications across industries, experts note.

“This case reinforces the importance of actually managing AI risks,” said Valence Howden, advisory fellow at Info-Tech Research Group. “If an AI-enabled model or ATS [Applicant Tracking System] is making decisions based on historical information, it can raise questions about whether bias in outcomes and datasets has been properly addressed.”

The case so far

Mobley v. Workday, Inc. alleges that Workday’s AI screening tools discriminate against job seekers based on age, race, and disability. The suit was filed in 2024 in the US District Court of California by Derek Mobley, a Black disabled man over 40, who claimed Workday’s algorithms continually screened him out as he applied for more than 100 positions on the platform.

The claims alleged discrimination prohibited by several US and California statutes: Race and sex under the Civil Rights Act of 1964 (Title VII); disability under the Americans with Disabilities Act of 1990 (ADA); age under the Age Discrimination in Employment Act of 1967 (ADEA); and race, gender, and age under California Fair Employment and Housing Act (FEHA).

Specifically, the suit centered around Workday’s use of automated, algorithm-driven tools for applicant screening. It alleged that these systems rely on historical data and statistical modeling that can make them susceptible to existing biases, even if protected characteristics like race, age, sex, or disability are not explicitly provided.

Bias may enter these systems in different ways, the plaintiffs argued, including via training data, model design, and evaluation criteria for candidate fit. The system could reproduce discriminatory outcomes by making correlations from data. For instance, years of experience on a resumé may indicate age; long employment gaps may infer a disability or caregiving responsibilities; educational and institutional affiliations could reflect race.

Workday has argued that it is not subject to liability under employment statutes because it does not qualify as the job applicants’ “employer.” But federal judges have allowed key parts of the lawsuit to move forward, ruling that Workday could potentially be treated as an employer’s “agent” for the purposes of anti-discrimination law.

The latest dispute centers on FEHA. According to legal sources, the California statute is among the strongest anti-discrimination laws in the US, in many cases providing broader protections than federal employment laws.

Workday asked the court to dismiss claims brought under California law, saying FEHA should not apply to the hiring decisions of out-of-state employers and applicants. The company’s lawyers argued that enforcing this would effectively allow California law to supersede that of other states, just because a company used their platform.

But Lin disagreed, saying FEHA does apply, and in fact, Workday is directly liable for its “own engagement in FEHA-regulated activities on the employer’s behalf.” Holding businesses liable for “their own discriminatory conduct” is within the scope and purposes of FEHA and consistent with public policy.

However, the issue is still to be decided; Lin did not indicate when she would release a final ruling.

Workday’s defense

A Workday spokesperson called the claims in the suit “false.”

“Workday’s AI recruiting tools don’t make hiring decisions and are designed with human oversight at their core,” the spokesperson told CIO. “Our technology looks only at job qualifications, not protected traits like race, age, or disability. We rigorously test our products as part of our responsible AI program to confirm our tools do not harm protected groups.”

Workday’s platform is meant to provide insights on how well a candidate’s qualifications match the requirements of a posted job, the company said. Those tools focus only on qualifications listed in a candidate’s application, which are compared to qualifications identified by the employer as important for the job.

Workday’s Chief Responsible AI Officer Kelly Trindel said its AI does not make employment decisions, automatically reject candidates, or determine who gets a job; further, she said, there is no evidence that the company’s tools result in harm to protected groups.

Trindel, who is former chief analyst of the Equal Employment Opportunity Commission (EEOC), leads a dedicated team composed of psychologists and PhD-level data scientists whose sole focus is to ensure that its AI is “responsible, fair, and ethical.” She said that the company’s AI systems undergo ongoing reviews throughout their lifecycle to help prevent unintended consequences, and Workday is “committed to accountability, transparency, and trust,” and invests “significant resources” into identifying and mitigating bias.

Further, she said, Workday has a company-wide commitment to ethical AI, and an independently-evaluated AI governance program based on standards from the National Institute of Standards and Technology (NIST) and the International Standards Organization (ISO).

“Workday builds AI to support people, not replace them, and this is of particular importance when it comes to hiring,” Trindel noted. Its platform is designed to help employers “manage high-volume processes more efficiently, surface relevant information, and reduce administrative work so teams can spend more time applying their expertise and judgement to hiring decisions.”

What this means for enterprise leaders

Workday isn’t alone in its legal challenges; other AI hiring tools are also being scrutinized over their methodologies, algorithms, and data-collecting practices. Eightfold, for one, is also facing a California class action lawsuit alleging that its tools unfairly rely on job candidates’ online data to predict whether they’d be a good fit for a position.

This means that enterprises, who are already feeling increased pressure to document hiring decisions, conduct AI bias audits, and maintain human oversight in recruitment and hiring, must be even more diligent in their vetting of AI tools.

Organizations must be actively defining how these recruitment tools should work, identifying bias in their algorithms, and setting up structures to test for bias across the tools’ decision-making logic, Info-Tech’s Howden advised.

“Validation of non-biased outcomes also needs to be active and ongoing, rather than a point-in-time exercise,” he said.

While Workday and others say human oversight is paramount, “it’s hard to incorporate humans into the process if the platform does the weeding out before humans have the ability to intervene,” Howden pointed out.

Discriminatory biases can exist in past hiring decisions, so it’s easy to forget that AI can “emulate and adapt those biases as part of its perspective,” he said. That includes how AI looks at language: Different cultures use different phrasing, and AI can capture that and use it to exclude candidates.

Ultimately, he called the case a “cautionary tale” illustrating how lightly some organizations have been treating AI risk. It also highlights the urgency involved in building out more advanced enterprise risk practices, “rather than relying on the limited capabilities they may have employed up until now.”

This article originally appeared on CIO.com.

For generations, technology export controls referred to the transfer of code to other countries. But that no longer works, as the latest Anthropic fight with the US Commerce Department makes clear. 

On Friday, Anthropic announced that it had received instructions from Commerce “to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Anthropic models will not be affected.” 

Technically, the Commerce letter doesn’t explicitly say that, but lawyers and consultants argue that, when combined with an earlier executive order declaring Anthropic a supply chain risk, that very well might be what it means.

What the Commerce letter says is that Anthropic needs a license to export Fable 5 and Mythos 5 (a “deemed export“), listing four circumstances in which that license would be required: “The sending or taking of the model out of the United States in any manner; The sending or taking of the model from one foreign country to another in any manner; Retransferring the model within a single foreign country; or the release of the model to a ‘foreign person’ in the United States or a foreign country.”

Restricts capabilities not code

Although moving a model has historically meant transferring the code, most experts argue that the definition has changed for all SaaS deployments, and could now be interpreted as referring to any access to the models.

“This is not just about data sovereignty anymore. It is about capability sovereignty, where governments want to control who has access to frontier AI capabilities, irrespective of who built it, where it is hosted, or who worked on it,” said Valence Howden, advisory fellow at Info-Tech Research Group.

“The reference to deemed exports is important, because traditionally that would apply to code, technology, or technical knowledge being transferred,” he said. “In this case, the thing crossing borders is not necessarily the model itself, but it is access to the capability. That is a significant shift, and signals the real intent behind the AI arms race. The focus is moving from controlling the technology to controlling access to the outcomes the technology can produce.”

Mark Rasch, a former federal prosecutor who specializes in legal technology issues, agreed. “I don’t need to have the code physically resident in order to take advantage of the capabilities of that code. Today, the location of the code is irrelevant.”

Practical challenges

There are two practical issues involved. The first is that a large part of Anthropic’s workforce is not US citizens, and some of them have direct access to the code for these models. 

But the potentially more daunting issue is that today it is difficult, if not impossible, to identify the citizenship of any AI user, which might force companies to assume that everyone might be unauthorized. 

In fact, said Yuri Goryunov, CIO of consulting firm Acceligence, “there is no way to check citizenship through an API call. Besides, three-quarters of Americans don’t have passports.”

Consultant Brian Levine, executive director of FormerGov, added that the issue will make life difficult for CIOs even if the Commerce position is viewed as dubious.

“Regardless of the strength of Commerce’s position, once it issues an ‘Is Informed’ letter, every unlicensed interaction with a foreign person becomes a potential violation, and the safest move is often to halt access until a licensing path exists,” he said.

This means that enterprise CIOs need to approach AI contracts with the knowledge that any government can now declare the product legally unavailable, with no notice. 

Sovereignty has climbed the stack

Howden said that this shift will force CIOs to strongly consider non-US AI models such as France’s Mistral or even China’s DeepSeek, “to reduce the concentration risk attached.” 

“There are hundreds out there that are very good. It is very easy to sit in the bubble of the four or five we know,” Howden said. “Enterprise CIOs think it’s a much more limited market than it really is.”

Sanchit Vir Gogia, chief analyst at Greyhound Research, said that enterprise executives now need to take the significant import rule changes into account when selecting AI models. 

“The harder truth is that sovereignty has climbed the stack,” he said. “The wall no longer stands around the database, it now stands around the intelligence layer itself. Under the export-control frame, it covers the release of controlled technology or code to a foreign person, including one standing inside the United States. The border follows the person, not the parcel. Code is part of the doctrine, but it is not the whole of it.”

He added, “the difficulty is that the cited rules speak of technology and code, while the letter reaches for the model itself. A hosted model hands the user no weights and no code. It hands them inference, and inference is a capability, not a file.”

Ultimately, Gogia said, Anthropic’s decision to cut off model access to everyone immediately “was a rational answer to an impossible instruction. A frontier model can now vanish for reasons unconnected to uptime, price or performance. The same models that help secure systems can be withdrawn at the moment defenders most need them.”

This article originally appeared on CIO.com.