DRAM and NAND. Behind these two acronyms lie essential components of modern consumer electronics, from smartphones and computers to gaming consoles. These two families of memory chips are experiencing runaway inflation, driving a wave of price increases — including those announced by Apple last week. The culprit is the breakneck expansion of infrastructure dedicated to generative AI.

While many had still hoped for some relief next year, that respite now appears unlikely to arrive before 2028. By then, new production lines should come online, helping to ease the supply shortage. Even so, few observers expect memory prices to collapse, let alone return to their early-2025 levels, at least not as long as the AI boom continues. High prices have become “the new normal,” a Lenovo executive said last week. “The era of cheap silicon is over,” Nothing founder Carl Pei warned earlier this year.

An oligopoly

DRAM, or dynamic random-access memory, provides temporary storage for the data used by a processor when running software and applications. It comes in three main varieties (GDDR, DDR and LPDDR) depending on requirements for speed and power consumption. NAND, by contrast, is flash memory that retains data even when a device is powered off. It is used in USB drives, memory cards and SSDs, which have largely replaced traditional hard drives in consumer electronics.

The DRAM market is dominated by an oligopoly of three companies — South Korea’s SK Hynix and Samsung, and US-based Micron — which account for more than 90% of global sales, according to Counterpoint estimates. China’s CXMT is the only other producer with meaningful volumes, though its chips are primarily destined for the domestic market. The NAND sector is more competitive. While the same three companies control around 60% of the market, another trio — US-based SanDisk, Japan’s Kioxia and China’s YMTC — share the remaining 40%.

High-bandwidth memory

The memory-chip industry is notorious for sharp price swings, regularly oscillating between periods of shortage and oversupply. At the end of 2022, for instance, prices collapsed after smartphone and, above all, PC sales declined following the pandemic-driven surge in demand. The following year, SK Hynix and Micron posted the largest operating losses in their histories, while Samsung’s profits plunged 77%, dragged down by its semiconductor division.

The current price surge truly began last autumn. It is being driven by soaring demand for high-bandwidth memory (HBM), which boosts processing speeds while reducing power consumption. Paired with GPUs and processors, HBM has become indispensable for training and running AI models. Demand for these chips is exploding, fueled both by the massive increase in computing power and by the growing number of HBM stacks integrated into each AI system.

Underinvestment

To meet demand, manufacturers have redirected part of their production capacity toward HBM, at the expense of conventional DRAM and NAND. According to TrendForce, around 70% of global memory production is now dedicated to HBM and other chips used in AI infrastructure. The result: DRAM prices have nearly quadrupled in just nine months. The trend is expected to continue. Analysts at Jefferies believe prices could double again by the end of this year before rising by another 50% in 2027.

SK Hynix, Samsung and Micron failed to anticipate such massive demand. Worse, all three companies underinvested in recent years following the last industry downturn. They now plan to expand capacity, but their new factories will not become operational until 2027 or 2028. And there is no guarantee that these facilities will not be devoted primarily, or even entirely, to AI-related memory products. The emergence of new categories such as robots, which will also require large amounts of memory, adds another layer of uncertainty.

RAMpocalypse

The situation is proving highly lucrative for memory giants, which are enjoying spectacular growth and record profits. But it is punishing the companies that buy DRAM and NAND. They are now forced to choose between raising prices, accepting lower margins or reducing memory capacity. Entire segments of the market could disappear. Smartphones priced below $100, for example, may no longer be economically viable. Budget brand CMF, a subsidiary of Nothing, has already canceled the launch of its next handset.

Beyond soaring prices, the industry also faces the threat of outright shortages, particularly in DRAM. According to analyst Ming-Chi Kuo, Apple may end up producing 10% to 20% fewer new iPhones than originally planned. Nikkei reports that Xiaomi now expects to ship 95 million smartphones in 2026, down from an initial forecast of 135 million and from 170 million units last year. According to IDC, this “RAMpocalypse” could trigger a historic contraction in the global smartphone market. PCs and tablets are also expected to suffer heavily.

Photo: Micron

The idea seems, to say the least, far-fetched. In a letter sent to Brussels, Austria’s digital minister suggested exploring Anthropic’s “strategic establishment” in Europe by offering the company “legal certainty, market access, capital and a set of values that suits [its own].”

Unrealistic as the proposal may be, it illustrates a major shift in the regulation of generative AI: after long opposing any form of intervention, the Trump administration is now asserting the authority to decide who can access the most advanced AI models.

“Trusted partners”

This new approach began to take shape in mid-June, when Washington barred foreign nationals from using Anthropic’s Mythos 5 model and its consumer version, Fable 5. The startup promptly disabled both systems, saying it was technically unable to comply with the order.

The policy hardened further on Friday, when OpenAI was forced to limit the rollout of its latest model, GPT-5.6, to around twenty “trusted partners”, almost certainly all American, whose identities were approved by the government.

In both cases, national security concerns have been cited. The models could allegedly be used by malicious actors to discover and exploit previously unknown software vulnerabilities. Both companies, however, have downplayed the risks.

Anthropic says it implemented safeguards in the consumer version of Mythos and argues that the dangers are “comparable” to those posed by other models on the market. OpenAI, for its part, maintains that GPT-5.6 is “unable to carry out autonomous, end-to-end attacks”.

A dramatic reversal

These episodes represent the culmination of Donald Trump’s reversal on AI policy. Upon returning to office in January 2025, the US president rescinded an executive order signed two years earlier that required AI companies to conduct safety tests on new models and share the results with the government.

He also relaxed environmental restrictions on the construction of data centers. The moves were applauded by Silicon Valley executives, who flocked to the White House to bow before its new occupant.

In recent months, however, the tone has begun to change — perhaps following the dispute with Anthropic over the Pentagon’s use of its models, or because of growing public concern, including among Republicans, about the rapid rise of AI.

In June, Trump signed an executive order introducing a 30-day government review of the most advanced models before their launch, though only on a voluntary basis. The final version of the measure had been watered down after lobbying by several technology companies.

Worse than regulation

The current situation appears more problematic than regulation itself. It demonstrates that, in the name of national security, the government can not only ban the export of certain models but also decide which domestic organizations and companies are allowed to use them.

More importantly, these decisions don’t rely on no clearly established legal framework, but on opaque criterias that may be arbitrary and are liable to shift with public opinion, the mood of political leaders, or their relationships with AI companies and their customers.

It is an ironic outcome for the industry’s leading players, which — with the notable exception of Anthropic — have spent the past three years fighting nearly every attempt to regulate AI. In Europe, they ultimately secured a softening of the AI Act. In the US, they helped derail almost every AI bill introduced at the state level.

The uncertainties created by Washington’s new doctrine may prove far more damaging than regulation itself. Regulation undoubtedly creates administrative burdens and compliance costs, but it also offers predictability.

Photo: The White House

Just three years ago, SK Hynix posted the largest loss in its history. Yet the memory-chip giant has now (briefly) overtaken Samsung to become South Korea’s most valuable listed company, ending its rival’s uninterrupted 26-year reign. The company is now looking to capitalize on its meteoric rise. On Wednesday, it confirmed plans for a dual listing on Nasdaq starting July 10. The offering could raise more than $29 billion, an unprecedented amount that will be used to finance the construction of new manufacturing facilities.

Profits expected to increase fivefold

Between its record losses and its likely triumphant arrival on Wall Street, one development reshaped the industry landscape: the explosive rise of generative AI. The boom has fueled both soaring demand, especially for high-bandwidth memory (HBM) chips, and an unprecedented surge in pricing. Already strong, the market’s momentum is accelerating at a remarkable pace. According to analyst forecasts, SK Hynix’s revenue could triple this year, while profits are expected to increase more than fivefold.

SK Hynix’s history has been marked by repeated upheavals and ownership changes. Founded in 1983 as Hyundai Electronics, the company initially focused on automotive components. During the Asian financial crisis, it merged in 1998 with LG’s semiconductor division, only to come close to bankruptcy three years later. Burdened by debt, it narrowly avoided being acquired by U.S. chipmaker Micron. By then operating independently under the name Hynix, the company was ultimately acquired in 2011 by telecom operator SK Telecom for $3 billion.

A winning bet

Barely fifteen years later, SK Hynix is now worth nearly $1.2 trillion on the stock market — nine times its value a year ago. The company is reaping the rewards of a strategic bet made in 2013: developing a new category of memory chips in partnership with US CPU maker AMD. Built on a radically different architecture, HBM chips provide significantly higher bandwidth, enabling faster processing speeds while reducing power consumption. They have become indispensable for both the training and inference of AI models.

According to estimates from Bank of America, the HBM market has expanded more than tenfold since 2023. It is expected to grow another 58% this year, reaching $55 billion. Only two other companies manufacture HBM chips: Samsung and Micron. Although rivals have partially narrowed the technological gap, SK Hynix remains the market leader. Research firm Counterpoint estimated its market share at 58% in the second quarter. Analysts also expect the company to supply 60% to 70% of the HBM4 memory chips paired with Nvidia’s next-generation Rubin graphics processors.

Memory prices set to surge in 2027

While demand for HBM chips is exploding, prices have remained relatively stable. Unlike DRAM and NAND memory, HBM pricing is negotiated through annual contracts. However, analysts believe this situation will be temporary. Bernstein forecasts significant inflation next year, with prices potentially rising by as much as 2.5 times — even for older-generation products. Such increases could drive SK Hynix’s profits to new heights. Last year, the company became South Korea’s most profitable business, narrowly surpassing Samsung.

AI infrastructure requires other types of memory as well, including LPDDR5X chips used alongside CPUs, which have become increasingly strategic. Demand continues to rise as computing power expands — and so do prices. According to Morgan Stanley, Nvidia’s next supercomputer, combining 72 GPUs and 36 CPUs, will contain roughly $2 million worth of memory, five times more than its predecessor. Research firm SemiAnalysis estimates that memory will account for 30% of cloud giants’ capital expenditures this year, up from just 13.5% in 2025.

A “disciplined” investment strategy

Like its competitors, SK Hynix initially redirected part of its production capacity toward AI-focused components. That strategy triggered a sharp rise in prices for DRAM and NAND memory used in smartphones, PCs, and gaming consoles. Prices quadrupled within nine months, contributing significantly to revenue growth. These products now generate even higher margins than HBM chips, to the point that the company is reportedly considering increasing production of conventional memory rather than accelerating AI-related output, according to the Kokean newspaper Chosun.

SK Hynix plans to expand manufacturing capacity, but its new plants will not begin operating until 2027 or 2028. While executives have signaled a “substantial” increase in investment, they insist on maintaining a “disciplined” approach, limiting capital expenditures to 30% of revenue. That caution reflects the cyclical nature of the memory industry, which regularly swings between shortages and oversupply. SK Hynix remains wary of a potential downturn should enthusiasm for AI begin to fade.

Photo: SK Hynix

Faced with the rise of generative AI, Getty Images has long favored litigation. On Sunday, however, the U.S. image licensing giant signaled a strategic shift by announcing a partnership with OpenAI.

Similar to the agreement it struck last year with startup Perplexity, the deal remains limited in scope: it does not cover AI model training, but will simply allow ChatGPT to display Getty photographs in its responses. Financial terms were not disclosed, yet the announcement alone was enough to send Getty’s stock soaring 90%.

Lawsuit against Stability AI

Founded in 1995, Getty Images offers a library containing hundreds of millions of photos and videos. The collection includes both news content, widely used by media organizations, and stock imagery, which accounts for the bulk of the company’s revenue.

That business is now under direct threat from the rapid advances in AI image-generation tools. Why pay hundreds of dollars for a stock photo when a comparable image can be generated by AI in seconds at a fraction of the cost?

As AI emerged, Getty filed a lawsuit against Stability AI in February 2023. At the time, Stability was the leading player in AI image generation. Getty alleges that the company used approximately 12 million images from its catalog without authorization or compensation to train its models.

The claim gained traction after some images generated by Stability’s tools appeared to contain Getty’s watermark — the logo visible on publicly accessible Getty photos. The British startup has not disputed the use of Getty images, although it secured an initial legal victory in the United Kingdom on procedural grounds.

Legal uncertainty

A trial is still expected in the United States. Its outcome will largely depend on how courts interpret the doctrine of fair use, which allows the use of copyrighted works to create new content with a “transformative” purpose.

Stability argues that it is operating within the law because its AI-generated images do not copy Getty’s photos but merely draw inspiration from them. In a separate case, however, a judge suggested that the fair-use defense may not apply when copyrighted material is used to create a “market substitute”—an argument Getty is likely to emphasize.

Beyond the legal uncertainty, another factor may explain why Getty has not targeted other AI companies: proving that its content was actually used for training can be difficult. As a result, the company appears to be favoring display agreements that allow it to capture a share of the value generated by AI systems.

That share remains relatively small, however, since the greatest value lies in training data rather than image display. Getty executives have not ruled out broader partnerships that could help monetize the company’s vast archive, even as competition from AI-generated imagery appears increasingly impossible to avoid.

What’s in it for OpenAI?

In this new environment, Getty is promoting its own AI tools, which allow customers to generate or edit visuals using content from Getty’s licensed library. Its pitch is straightforward: legally protected content that does not rely on images allegedly used without authorization for model training, thereby reducing potential legal risks for users.

Whether that promise will prove compelling remains unclear. Getty is also betting on its acquisition of rival Shutterstock, a deal expected to double revenue while generating significant cost synergies.

For OpenAI, the benefits of the partnership appear more limited. The company behind ChatGPT did not even issue a press release announcing the agreement. ChatGPT can already display images sourced from across the internet in response to certain queries.

While access to Getty’s library could improve image quality, the strategic significance appears modest. The deal may simply represent a first step toward a broader agreement covering AI training data. Alternatively, it could be little more than a way for OpenAI to shield itself from potential legal action by Getty.

Photo: Getty Images

“Unfortunately, price increases are unavoidable”. Before stepping down on September 1, Tim Cook delivered unwelcome news to Apple customers: the next generation of iPhones will most likely be more expensive than their predecessors. The reason is the sharp rise in memory chip prices, driven by booming demand from generative artificial intelligence. “The situation has become unsustainable,” the Apple CEO told The Wall Street Journal, without providing further details on the timing or the products affected.

So far, Apple has chosen to absorb the higher costs, relying on its strong margins to avoid hurting demand. In the first quarter, iPhone sales even posted a slight increase while the broader smartphone market returned to decline, weighed down by rising prices in the entry-level and mid-range segments. The launch of new models in September, however, provides an ideal opportunity to adjust prices, not only to pass on at least part of the costs already incurred, but also to prepare for further increases expected in the coming months.

Prices Up fourfold

The situation is unlikely to improve before mid-2027 at the earliest. At the heart of the disruption is the rapid expansion of AI infrastructure, which requires enormous amounts of memory, particularly high-bandwidth memory (HBM) chips. To meet this demand, the industry’s three dominant players — South Korea’s SK Hynix and Samsung, and US’ Micron — have redirected part of their manufacturing capacity toward these more profitable products, at the expense of DRAM and NAND memory used in smartphones, PCs, tablets and gaming consoles.

As a result, prices for those chips have soared. According to estimates from research firm TrendForce, memory prices have nearly quadrupled in just nine months. While manufacturers plan to expand capacity, bringing new production lines online takes time. More importantly, most, if not all, of those investments are aimed at AI-focused memory products. Morgan Stanley analysts expect DRAM production for AI applications to rise by 30% by 2027, while output destined for consumer electronics could decline by 15%.

A $144 increase in costs?

According to estimates from research firm TechInsights cited by The Wall Street Journal, the DRAM and NAND memory components in the iPhone 17 Pro cost Apple around $52 at launch last September. That represented roughly 9% of the device’s total component cost — a particularly low figure by industry standards. For its successor, expected this fall, the bill could jump to $196, or 22% of the total, assuming other component costs remain broadly unchanged. DRAM would become the most expensive component in the smartphone, overtaking both the processor and camera module.

While these figures remain estimates based on market prices rather than Apple’s negotiated contracts, they illustrate the dilemma facing the company. Without a retail price increase — and excluding potential rises in other component costs — the iPhone’s gross margin would fall from 47% in September 2025 to 34% a year later. To maintain the same profitability, Apple would need to raise the price of the device by roughly $270, a premium that appears too large to pass on entirely to consumers.

No more safety cushion

More likely, the Cupertino-based giant will opt for a price increase of between $100 and $200 across its new iPhone lineup. Such an adjustment would help limit margin erosion in the short term without fully offsetting the expected rise in memory costs over subsequent quarters. The impact could therefore be twofold: lower profit per device sold and weaker sales volumes, including for older models whose prices may no longer decline as they typically do.

Until now, Apple has been relatively shielded from rising memory costs. During the first three months of the year, its hardware gross margin slipped only marginally thanks to inventory built up in advance. But that buffer will not extend beyond June, Tim Cook warned in April. To mitigate the impact, Apple’s CEO said he is prepared to use the company’s cash reserves, potentially by committing to long-term supply contracts similar to those signed by major AI companies. The decision, however, will ultimately fall to his successor, John Ternus.

Photo: Apple

DeepSeek has taken its time before capitalizing on the success of its generative AI models. Eighteen months after bursting onto the global stage in what was dubbed AI’s “Sputnik moment,” the Chinese startup has completed the first fundraising round in its history, securing 50 billion yuan (€6.4 billion). The deal, based on a $50 billion valuation, will finance the company’s long-term research efforts, founder Liang Wenfeng reportedly told prospective investors, according to Bloomberg.

Reflecting DeepSeek’s strategic importance for China, the funding round includes no U.S. venture capital firms. It does, however, feature the National AI Investment Fund, launched by Beijing last year to support domestic AI champions. Although the state-backed fund is contributing only 1 billion yuan, it will enjoy voting rights that other shareholders will not. Most of the capital comes from High Flyer, the hedge fund that owns DeepSeek, while the remainder is being provided by major Chinese companies, including Tencent, the owner of messaging platform WeChat.

Just $5.6 million?

Founded three years ago, DeepSeek emerged from relative obscurity in January 2025 with the release of its V3 and R1 models, which rivaled offerings from OpenAI and Anthropic. The lab lacked access to Nvidia’s most advanced graphics processing units (GPUs), whose exports to China are prohibited. Its researchers compensated for this handicap by developing new training techniques. That narrative, however, has been questioned. OpenAI suspects the company relied on its own models during development, while US authorities have accused it of illegally acquiring restricted Nvidia chips.

DeepSeek also claims that developing V3 cost just $5.6 million — a tiny fraction of the hundreds of millions, or even billions, spent by leading US AI firms. The figure, which cannot be independently verified, triggered panic on Wall Street: Nvidia’s stock plunged 17% in a single trading session. Although the competitive landscape has not shifted as dramatically as some predicted, Chinese AI companies have continued to gain ground against the closed models offered by Western giants, including in the US, thanks to aggressively low pricing.

New models

Following its debut, DeepSeek’s next generation of models took much longer than expected to arrive. Originally scheduled for release in May 2025, they were postponed several times. The delays reportedly stemmed in part from the startup’s efforts to use AI chips designed by Huawei, which are less powerful. In April, DeepSeek finally unveiled its V4 family, claiming performance on par with the most advanced models from OpenAI, Anthropic and Google — although those benchmarks have since been surpassed by newer iterations. This time, the company did not disclose training costs.

Questions remain here as well. Earlier this year, The Information reported that DeepSeek had obtained Nvidia’s latest Blackwell GPUs through shell data centers located overseas and used as intermediary hubs. The startup insists it relied exclusively on Huawei’s newest AI chips. At the same time, it says it has continued refining its training methods. In early January, the company published a research paper describing a new approach designed to further reduce computing requirements — and therefore costs.

Rock-bottom prices

Eighteen months after DeepSeek’s breakthrough, its true weight within the AI ecosystem remains difficult to assess, largely because no financial data has leaked to the press. In a report published in January, Microsoft estimated that it held less than 5% market share in the US and major European countries, compared with 89% in China. Since then, the rapid adoption of AI coding tools and autonomous agents may have further strengthened the Chinese company’s position. The latest example: Microsoft is reportedly considering offering DeepSeek’s V4 model through its new agentic platform, Copilot Cowork.

To capitalize on this momentum, DeepSeek has slashed the price of its APIs — the interfaces that allow developers to integrate AI into applications. One million tokens, the standard unit used to measure model usage, costs just 88 cents with V4, compared with $25 for Anthropic’s latest model and $30 for OpenAI’s. The aggressive pricing strategy is intended to attract companies eager to cut AI costs while helping DeepSeek fend off growing competition from other Chinese players, including Alibaba and Moonshot.

Photo: DeepSeek

Nearly $20 billion in operating losses last year. OpenAI’s financial results — first obtained by author Ed Zitron, a vocal critic of the generative AI business model, and later confirmed by the Financial Times — are unprecedented. Yet the creator of ChatGPT is reportedly considering slashing prices for enterprise customers in an aggressive bid to gain market share from its main rival, Anthropic, even if doing so would mechanically deepen its already enormous deficit.

Spending spirals higher

OpenAI’s financial performance tells two very different stories. On the one hand, the company led by Sam Altman is growing rapidly: revenue reached $13.1 billion in 2025, up from just $3.7 billion a year earlier. On the other hand, expenses are rising just as quickly, climbing to $34 billion from $12.5 billion in 2024. Of that total, $19 billion was spent on research and development, primarily to train increasingly compute-intensive AI models, including $10.6 billion paid to Microsoft.

On the positive side, OpenAI can point to an improvement in its gross margin. In 2024, it stood at just 29%, meaning the company spent 71 cents on inference—the process of generating text, images or code — for every dollar of revenue. Last year, that margin rose to 43%. Even so, it remains far too low to cover the company’s other expenses, including model training as well as sales and marketing costs, which increased fivefold over the past year.

Pressure on margins

As a result, OpenAI’s operating losses more than doubled last year, rising from $8.8 billion to $20.1 billion. Its net loss was even more striking, reaching $38.5 billion compared with $5 billion in 2024. Much of that figure, however, reflects accounting adjustments related to the company’s legal restructuring, finalized at the end of October. Its deficit is also likely to widen significantly this year, this time because of accounting changes tied to its planned initial public offering, which could come as early as September.

OpenAI’s operating losses are driven largely by two business segments with negative margins. First, its free user base of nearly 900 million people generates no revenue while still incurring inference costs, pending the hoped-for expansion of advertising income. Second, heavy users — particularly those relying on its coding tool Codex — often generate costs that far exceed the value of their monthly subscription fees, at least until the company potentially shifts to usage-based pricing.

Undercutting Anthropic

Despite these staggering losses, OpenAI may become even more aggressive. According to The Wall Street Journal, the company is considering cutting the price of its APIs, the interfaces that allow developers and businesses to integrate AI models into applications and workflows. Access to its latest GPT-5.5 model currently costs $30 per million tokens—the units used to measure AI consumption. That is twice the price charged before the launch of what was then its flagship model.

More importantly, GPT-5.5 costs $5 more per million tokens than the latest version of Anthropic’s Claude, despite being regarded as less capable on certain tasks, particularly code generation—the fastest-growing segment of the AI market. By undercutting its rival on pricing, OpenAI hopes to expand its market share at a time when AI costs are becoming a growing concern for corporate customers. But such a strategy would squeeze margins on its API business, reducing its ability to absorb the company’s overall spending.

Chinese open-source models

As software development tools continue to gain traction—and with autonomous AI agents expected to drive another wave of adoption—the bill for companies using APIs is soaring while becoming increasingly difficult to predict. Token consumption has emerged as a strategic issue. And “tokenmaxxing,” an internal policy aimed at encouraging employees to maximize AI usage, is no longer in vogue. Illustrating the shift, Meta recently instructed staff not to “use AI tools simply for the sake of it,” while Uber has introduced monthly spending caps.

Faced with rising costs and growing awareness among enterprise customers, generative AI providers face a twofold risk. Existing clients may scale back usage, while others could increasingly turn to cheaper alternatives, particularly open-source Chinese models developed by DeepSeek, Alibaba and Moonshot AI, where a million tokens costs only a few dollars. OpenAI will never be able to match those prices. But by lowering its own rates, it may be able to limit customer migration to low-cost competitors for less demanding workloads.

Read more:
– Why OpenAI wants to turn ChatGPT into a super app
– Heavy losses and weak revenue: xAI’s brutal numbers

Photo: OpenAI

Behind the fanfare surrounding its latest launch may lie a genuine commercial revolution in the generative AI market. On Tuesday, Anthropic released a public — albeit limited — version of its Mythos model, which had until now remained restricted for safety reasons. Beyond benchmark results, which rank it well ahead of competing systems, the launch comes with a significant change in pricing: starting on June 23, access to the model will no longer be included in paid subscriptions, including the company’s $200-per-month plan.

The model, called Fable 5, will instead be subject to usage-based billing on top of the monthly subscription, with charges calculated according to the number of tokens consumed — the units of text used to measure AI model usage. The era of fixed-price subscriptions constrained only by usage caps may therefore be coming to an end. Bills could become substantially higher, especially given Anthropic’s pricing: $50 per million tokens, compared with $25 for its previous flagship model and $30 for OpenAI’s GPT-5.5.

Only a temporary change?

Usage-based pricing is not new. It is already standard for APIs, which allow AI models to be integrated into applications and business workflows. In April, Anthropic also shifted customers of its enterprise offering to this billing model. However, this is the first time it has been extended to all Claude users. In doing so, Anthropic is following the path opened by GitHub, Microsoft’s developer platform, which recently adopted a similar pricing structure for its Copilot coding assistant.

The company insists the move may only be temporary, saying it “aims” to bring Fable 5 back into subscription plans “as quickly as [it] can”, though it stops short of making a formal commitment. According to Anthropic, the shift is driven solely by technical constraints: it does not yet have enough computing capacity to handle what it expects will be both “very high” and “difficult to predict.” demand. Having long lagged behind on infrastructure, the company is now investing aggressively to expand its computing resources.

The rise of agentic AI

Whether Anthropic ultimately follows through on that promise remains to be seen. In the meantime, the decision marks another step toward a new economic model for generative AI. The underlying challenge is well known: the high cost of inference — the process of running AI models to generate text, images, or code. As a result, profit margins remain slim and can even turn sharply negative for heavy users, making it difficult to cover both model training expenses and the massive investments required for computing infrastructure.

Until now, leading AI companies have largely absorbed these costs. Doing so was both a strategic choice and a necessity to drive adoption and reach critical scale. But the economics have deteriorated further with the emergence of autonomous AI agents capable of operating continuously, particularly in software development. AI tools no longer simply assist developers by suggesting or correcting code; they can now run “long, multi-step coding sessions” as GitHub noted when explaining its own pricing changes.

AI billed like electricity?

The losses associated with certain subscription plans have therefore widened significantly. While no company has released specific figures, screenshots posted on Reddit by GitHub Copilot customers provide an idea of the approximate cost. They suggest that some users would have incurred monthly bills running into the hundreds or even thousands of dollars under the new usage-based model. Similar estimates have been made for heavy Claude Code users who still benefit from Anthropic’s subscription plans.

Under such conditions, the traditional subscription model becomes increasingly difficult to sustain for certain use cases. Even OpenAI, despite its aggressive push to catch up in AI coding tools, has begun charging large enterprise customers based on actual consumption for its Codex product. In March, OpenAI CEO Sam Altman hinted at a broader shift in the industry’s business model. The future, he suggested, will not revolve around selling subscriptions but rather “selling tokens.” In the long run, he argued, AI could become “a public utility comparable to electricity or water,” with users paying according to the amount they consume.

Photo: Unsplash / Planet Volumes

At OpenAI, one super app may be hiding another. Last autumn, the maker of ChatGPT set out to follow in the footsteps of WeChat, the Chinese messaging platform that popularized the concept of mini apps. The idea was straightforward: open up its flagship chatbot to third-party services — with Spotify, Booking, and Expedia among the first announced partners — to make it the central gateway to users’ digital lives and reinforce its indispensable status in the face of growing competition. In practice, however, that vision never fully materialized.

Eight months later, OpenAI is preparing to launch a new super app, according to the Financial Times, confirming reports that first surfaced in March. A sign of the upheaval sweeping the generative AI market, the new platform would combine ChatGPT’s conversational interface with code-generation tools and AI agents. Those have become strategic priorities for the company as it seeks to drive revenue growth at a time when user growth is slowing and conversion rates to paid subscriptions remain limited.

Personal agents

OpenAI’s future super app is intended to mark the beginning of a paradigm shift. “Chat is dead,” one employee told the British newspaper. The AI pioneer believes the industry is moving toward personal agents capable of autonomously carrying out tasks on an ongoing basis, inspired by the OpenClaw phenomenon, whose founder it recently hired. Access to the technology remains limited because installation and configuration still require technical expertise. Google and Microsoft have nevertheless unveiled consumer-oriented AI agents in recent weeks.

Already the undisputed leader in the consumer market, with nearly one billion weekly active users, OpenAI now aims to “give agentic capabilities to everyone,” co-CEO Fidji Simo wrote in an internal memo revealed by The Wall Street Journal in March. The executive, who has since temporarily stepped back from her role for health reasons, argued that integrating these new agents directly into ChatGPT would allow the company to capitalize on its massive audience and strong brand recognition to position itself at the center of the next AI revolution.

Accelerating codex adoption

In the meantime, the super app is also expected to give a boost to Codex, OpenAI’s programming tool designed to compete with Anthropic’s Claude Code. The company says adoption has been strong, with its user base increasing sixfold since the launch of a dedicated desktop application in February. In April, OpenAI introduced new agentic features to expand the tool’s capabilities. “Codex is becoming a productivity tool for everyone,” the company says, citing use cases such as report writing, data analysis, and task automation.

By combining Codex with ChatGPT, OpenAI hopes to showcase these new capabilities to mainstream users and persuade a portion of its free user base to upgrade to a paid subscription. Reflecting that ambition, the company has placed both platforms under the leadership of a single executive. At the same time, it has shelved several side projects, including its Sora video model and its social network initiative, to concentrate resources on its new priorities — bringing to an end what Fidji Simo described as a strategy of “trying to do everything at once.”

Photo: OpenAI

AI-generated song remixes, previously off-limits, are making their way onto Spotify in a big way. Last week, the Swedish music streaming platform announced a partnership with Universal Music, allowing users to create tracks based on the major label’s catalog. Spotify presented the move during an investor event as the first “legal” offering that will compensate artists and songwriters, immediately setting it apart from services offered by U.S. startups Udio and Suno.

The feature will be offered as a paid add-on to existing subscriptions, although all users will be able to listen to AI-generated tracks. Neither pricing nor a launch date has been disclosed. Artists represented by Universal — including Taylor Swift, Drake, Adele, Elton John and U2 — will have to give their consent before their songs can be remixed. While the compensation model has not been detailed, it will likely follow the current royalty structure. Spotify said it is not developing its own AI models, but relying instead on external providers.

39% of songs generated by AI

The rise of generative AI was initially viewed as a threat by the music industry. Streaming platforms are now flooded with songs created using Suno, Udio and other tools capable of replicating voices, writing lyrics and composing melodies. On Deezer, for example, AI-generated tracks accounted for 39% of new songs added to the catalog in January, up from just 10% at the start of 2025. They were also capturing a growing share of listeners’ attention: 3% of total listening time, versus only 0.5% six months earlier, according to the French platform.

The threat is primarily financial. Songs generated in minutes from a simple prompt could compete directly with tracks performed by real artists, reducing the royalties musicians receive from streaming services. Record labels initially responded by turning to the courts. But their lawsuits have run into legal gray areas. The majors have therefore adjusted their strategy: rather than fighting what increasingly appears inevitable, they are seeking to shape the shift and capture a significant share of the value it creates.

Toward a “superfan” offering?

Last year, Universal and Warner Music both struck deals with startup Udio — though negotiations with Suno have reportedly proven more difficult. Spotify’s agreement follows a similar logic. It serves two purposes. First, it enables artists and record labels to generate additional revenue, whereas they currently receive nothing when AI systems draw inspiration from their songs. Second, it creates closed ecosystems: tracks will not be downloadable and therefore cannot be uploaded to rival platforms to generate royalties at rights holders’ expense.

For Spotify, the deal is not only a way to embrace the rise of AI in music and prevent new competitors from emerging, but also an opportunity to monetize the trend. The streaming leader, which aims to reach one billion listeners by 2030, up from 761 million today, hopes to unlock a new growth driver and raise average revenue per user alongside the multiple price increases introduced over the past three years. It may also mark the first concrete step toward the long-promised but still-unreleased “super premium” tier aimed at devoted superfans.

Photo: Spotify

When Elon Musk launched xAI in late 2023, he claimed the company could dethrone OpenAI, then the undisputed leader in generative artificial intelligence. The billionaire now admits that his start-up competes with the maker of ChatGPT mostly on paper. xAI holds “a very small market share,” Musk acknowledged in late April while testifying in an Oakland courtroom during the lawsuit he filed — and lost — against his rival’s chief executive, Sam Altman. That reality is reflected in the financial data made public last week as part of SpaceX’s IPO process.

Officially folded into the space company following an unlikely merger completed in February, xAI generated $3.2 billion in revenue in 2025. But most of that figure came from X, the former Twitter platform Musk acquired in late 2022 before selling it to his AI company two and a half years later. The social network generated $1.8 billion in advertising revenue, up from the previous year but still far below pre-acquisition levels. Added to that were revenues from its 4.4 million premium subscribers.

Weak revenue…

While SpaceX does not disclose revenue for Grok, its AI chatbot, the figure can nevertheless be estimated. On one hand, the company says it has 1.9 million paying subscribers. On the other, it cites an average monthly subscription price of $12, reflecting lower pricing in some countries. That translates into annualized revenue of roughly $270 million — half the target Elon Musk had reportedly set, according to Bloomberg. By comparison, OpenAI said in April that it had reached $24 billion in annualized revenue, while Anthropic reported $30 billion.

The estimate remains imperfect — it does not account for API revenue, nor for the share of X premium subscribers who signed up to access certain Grok features. Still, it shows that the start-up remains far from the ambitions of its founder, who invested tens of billions of dollars to build two supercomputers, dubbed Colossus, intended to support the expected growth in Grok usage. Those computing capacities are now underused enough that SpaceX is leasing them to Anthropic.

… and heavy losses

The gap between spending and revenue is clearly reflected in the company’s financial results. Last year, xAI posted an operating loss of $6.4 billion. The deficit directly tied to AI is likely even larger, since X posted a positive operating result in 2024, according to The Wall Street Journal. The company spent more than $5 billion on research and development alone, a budget that increased more than fourfold in a year and consisted largely of purchases of graphics processing units for its data centers, depreciated over several years.

In recent months, the start-up has also faced a wave of departures, including among its co-founders. “xAI was not built right first time around, so is being rebuilt from the foundations up,” Elon Musk said in March. The billionaire is particularly eager to accelerate development in AI coding tools, the industry’s latest battleground. Lagging far behind competitors, SpaceX announced in early May a deal to acquire Cursor, a start-up specializing in the field, for $60 billion. But it may already be too little, too late.

Netflix calls it a one-off. But it’s hard not to see the early signs of a major strategic shift. Scheduled for release next February, Narnia: The Magician’s Nephew will get a wide theatrical rollout across many countries. It’s a first for a feature film financed by the US streaming giant. Another notable change: the film will only be added to the platform’s catalog 49 days later, an unprecedented delay.

The timing of the announcement may not be accidental. At the end of February, Netflix walked away from a potential acquisition of Warner Bros Discovery, owner of the iconic studios and premium network HBO, after refusing to match a higher bid from Paramount. Executives had viewed the deal as a key growth lever amid a saturated market. Against this backdrop, Narnia could serve as a test case for a company historically driven by data, helping assess both incremental revenue and the impact on subscriber numbers.

$745 million at the box office

With an estimated $200 million budget and a star-studded cast including Meryl Streep and Daniel Craig, Narmia is no ordinary production for the Los Gatos-based group. It marks the first time the company adapts an already successful franchise, previously brought to the screen three times. At the request of director Greta Gerwig, fresh off the box-office triumph of Barbie, the film was already set for special treatment, with a two-week exclusive run across roughly 1,800 IMAX screens.

Extending that window was necessary to persuade major theater chains to screen the film without fearing that audiences would simply wait for its streaming release. The commercial upside appears significant. Released between 2005 and 2010, the first three adaptations of C. S. Lewis’s work grossed between $415 million and $745 million worldwide — roughly half of which went to theater operators. One unknown remains: marketing spend, often critical to a film’s success.

Oscar eligibility

Until now, Netflix had consistently refused wide theatrical releases for its productions. Instead, it limited screenings to a few hundred theaters at most, typically for just one week. The sole objective: qualifying films for the Oscars, in hopes of finally winning Best Picture — a prize that has eluded it despite eight nominations. This approach has long drawn criticism from producers and directors, some of whom refused to work with the platform.

Executives have often acknowledged that this strategy comes at a financial cost. It also runs counter to industry norms. Despite its push into streaming, Disney has maintained theatrical releases. Likewise, Apple and Amazon, following the acquisition of MGM, continue to release films in cinemas to generate additional revenue and reach wider audiences. Still committed to its direct-to-consumer model, Netflix argues it comes out ahead in the long run through subscriber growth.

Polishing its brand image

Its stance had already softened during the Warner deal attempt: to reassure Hollywood, Netflix pledged to preserve a 45-day theatrical window. Such a pivot wouldn’t be unprecedented. In recent years, the company has reversed several core principles. In late 2022, it embraced advertising, launching a cheaper tier to expand its subscriber base. It is also increasingly investing in live sports, gradually aligning itself with competitors’ strategies.

Beyond financial considerations, theaters could help enhance the perception of its in-house productions. Despite a high output — up to one film per week — Netflix has never replicated the cultural impact of its flagship series like House of Cards, Stranger Things, or Squid Game. Not for lack of investment, but largely due to its reliance on big-name stars to attract subscribers. Over the past three years, however, the company has shifted course, claiming to prioritize quality over quantity.

Photo: Unsplash / Charles Deluvio

There will be no courtroom showdown between OpenAI and Microsoft. Last week, the two partners formalized a new “long-term” agreement, removing the threat of legal action from the Redmond-based giant. This amicable split notably marks the end of their exclusive relationship — one that had gradually become a drag on the commercial expansion of the creator of ChatGPT, which in recent months has been caught up, and in some areas overtaken, by its main rival Anthropic. Just one day after, OpenAI announced a deal with Amazon to distribute its models on its cloud platform AWS, the global market leader.

In return, Microsoft secured the removal of a clause that would have allowed OpenAI to cut off its access to its technologies once artificial general intelligence, capable of learning any task, was achieved. The company will now retain its license through at least 2032. Another compromise concerns financial terms: Microsoft will still receive 20% of its partner’s revenue, but these payments are now time-limited — through 2030 — and capped. It will also no longer have to pay to offer OpenAI models on its Azure cloud.

A wake-up call for Microsoft

This marks the second time in six months that the two companies have revised the terms of their 2019 agreement — an alliance that had long been mutually beneficial. OpenAI gained the funding ($15 billion in total) and computing power needed to train its large language models. Microsoft, for its part, secured a share of revenue and potentially profits while gaining exclusive distribution of APIs on Azure, a major commercial advantage over AWS and Google Cloud.

The turning point came in November 2023, when OpenAI’s board abruptly dismissed Sam Altman, accusing him of racing too aggressively toward artificial general intelligence while overlooking risks. Although he was quickly reinstated, the episode served as a wake-up call for Microsoft, which realized its risky dependence on the startup. The company subsequently launched a new AI division and declined to invest further in OpenAI, forcing it to seek new funding sources.

End of exclusivity for training

Without much difficulty, the startup raised nearly $50 billion, notably from Japan’s SoftBank. But it also committed to its new investors to abandon its capped-profit status — failing which it would have had to repay those funds. This transformation required Microsoft’s approval. Protracted negotiations fueled growing tensions between the two partners, until a deal was finally reached in late 2025, allowing Microsoft to secure a 27% stake in the new legal entity.

The agreement also introduced initial flexibility for OpenAI, particularly in training and inference. Originally, the company was restricted to Azure servers, a limitation that constrained its ability to scale computing capacity to meet its needs. This exclusivity was partially lifted in 2024, though Microsoft retained veto rights. Freed from this constraint, OpenAI signed a $38 billion, seven-year contract with Amazon the following week.

Responding to Anthropic

Earlier this year, OpenAI deepened its ties with the e-commerce giant, which pledged up to $50 billion in investment, including $15 billion upfront. In return, OpenAI plans to spend an additional $100 billion on AWS and use multiple generations of Trainium chips designed by Amazon. The Seattle-based group also secured exclusive distribution rights for a new platform dedicated to deploying AI agents—a provision that sparked strong objections from Microsoft, which viewed it as violating its exclusivity clauses with OpenAI.

The agreement announced last week puts an end to that potential legal clash. More importantly, it opens the door to new distribution channels — a partnership with Google now appears likely. OpenAI is effectively responding to the rise of Anthropic, currently the only major AI player present across all three cloud giants. The stakes are high, especially as some internal targets have reportedly been missed, according to The Wall Street Journal. So much so that the company’s CFO is said to be concerned about its ability to finance the massive investments required to scale computing power.

Photo: Microsoft

Brussels is stepping up efforts to inject competition into online search. Last week, the European Commission said it intends to require Google to share its search data with third-party engines, as well as with generative AI services such as ChatGPT. The precise terms of the measure are now subject to public consultation, with a final decision expected by the end of July. Google, which has denounced an “overreach,” is likely to challenge the move before European courts.

The initiative is part of the Digital Markets Act, which came into force two years ago to boost competition in digital markets. Since then, several actions have been launched against Google, accused of failing to implement sufficient changes to comply. The issue of search data is being handled through a technical “specification” procedure, which does not involve fines but allows regulators to impose corrective measures if no agreement is reached.

At least five years of data sharing

Several types of information could be made available: queries, links and their ranking in results, as well as clicks. User-related data, such as geolocation, language, and device, would also have to be shared, once anonymized. These elements could help search engines and AI services better understand user intent and more accurately identify relevant pages—ultimately improving their results and narrowing the gap with Google.

The Commission does not plan to impose eligibility criteria: any search engine could request access. This would also apply to AI chatbots that rely on online searches to generate answers. To obtain the data, companies would need to sign a contract with Google, including restrictions on how the information can be used. The data would then be shared on a continuous basis for at least five years, giving competitors time to “optimize their services and challenge Google’s dominant position.”

Search index excluded

Access to the data will not be free. However, Google will have to offer a “reasonable” price, covering only its costs. It may apply a regulated margin only to companies operating at “very large scale” — those with more than 45 million users in Europe or a market capitalization above $75 billion. This would notably include Bing and ChatGPT, but not Perplexity AI. Pricing must also be non-discriminatory, meaning identical for all players.

Crucially, the obligations will not extend to Google’s search index, its database of web pages crawled by its bots. This remains one of the company’s key competitive advantages, enabling more relevant results. In the US, however, following an antitrust ruling, tje company could be forced to open this index — on a one-off basis rather than continuously — to rival search engines and AI platforms.

Ongoing compliance battles

This case is only one front in Google’s broader disputes with the Commission under the DMA, on top of previous fines now under appeal. A second “specification” procedure is underway, which could force the company to grant rivals of its chatbot Gemini equal access to features on its Android mobile system. More importantly, Google has faced a dual non-compliance investigation for the past year, which could result in substantial fines of up to 10% of its global revenue.

Brussels argues that the company gives preferential treatment to its own services, such as hotel booking and price comparison, systematically favoring them in search results over competitors. It also accuses Google of maintaining restrictions on “steering” (redirecting users from an app to a website to complete a purchase), and of charging commissions that “exceed what is justified” for such transactions. Google has since proposed adjustments, which are currently under review by the Commission.

Photo: Unsplash / Shutter Speed

John Ternus is one of Apple’s least visible senior executives. Yet he is set to become its next chief executive. On Monday, after months of speculation, Apple officially confirmed his appointment. He will replace Tim Cook, who has led the company since 2011 and will transition to executive chairman. The handover is scheduled for September 1 and comes at a pivotal moment: lagging in generative AI, Apple must now prove it can stay competitive.

John Ternus joined Apple in 2001 and currently heads hardware engineering, overseeing the design of the iPhone, Mac and other flagship products. According to Bloomberg, his responsibilities were expanded in late 2025 to include oversight of both hardware and software design — a strategic role once held by Jony Ive and later by Jeff Williams, who was once touted as the future CEO before his retirement last summer. John Ternus has since emerged as the front-runner to succeed Tim Cook.

Questions over innovation

Recruited in 1998 by Steve Jobs, shortly after Jobs’ return to Apple, Tim Cook is not a technologist by training. A former IBM executive, he built his reputation as an operations specialist. He notably overhauled Apple’s supply chain in China, enabling the company to scale production to hundreds of millions of devices annually. Appointed chief operating officer in 2007, Tim Cook served multiple times as interim CEO before officially taking over in 2011, succeeding Steve Jobs, who died two months later.

For 15 years, its tenure has unfolded in its precedessor’s shadow. Critics have pointed to a perceived lack of breakthrough innovation, highlighted by the commercial failure of the Apple Vision Pro and the abandonment of Apple’s autonomous electric car project. Yet Tim Cook has sustained remarkable growth, notably by expanding into services. Since 2011, revenue and profits have quadrupled, while market capitalization has increased tenfold.

His legacy may ultimately hinge on the coming months. Late to generative AI, Apple postponed the rollout of its revamped Siri, first unveiled in June 2024. In January, the company formalized a partnership with Google to integrate a customized version of Gemini. While this move signals a gap in Apple’s in-house capabilities, it is also pragmatic—buying time as the company develops its own models. The delayed Siri overhaul is now expected later this year.

Preparing for a post-smartphone era

John Ternus represents a sharp contrast to Tim Cook. An engineer by training, he has risen through the ranks of Apple’s hardware teams. In 2020, he took charge of iPhone development, before overseeing the company’s entire product lineup a year later. Though less public-facing than some peers, he has become increasingly prominent at Apple keynotes — most recently unveiling the new iPhone Air last September.

Although he was the front-runner, John Ternus wasn’t the only one in the running. The high-profile Craig Federighi, head of software engineering, and Eddy Cue, head of services, were also among the credible candidates. So were Greg Joswiak, who oversees marketing, and Deirdre O’Brien, head of Apple Stores. According to Bloomberg, the most serious alternative was chief operating officer Sabih Khan, appointed last summer — a role previously held by both Tim Cook and Jeff Williams.

Ultimately, Apple’s decision reflects more than a choice between internal candidates. It signals a strategic direction. On one side stood an operations-driven growth model, embodied by Sabih Khan and aligned with Tim Cook’s legacy. On the other, a renewed focus on technological innovation, represented by John Ternus. That choice may have been shaped by the next disruption Apple must confront: the future of devices themselves. Several players, including OpenAI — which has recruited Jony Ive — are already exploring what comes after the smartphone.

Photo: Apple

After two and a half years of steady growth, the smartphone market is heading sharply back into the red. First-quarter shipments to distributors fell 4.1%, according to IDC estimates. The primary culprit: surging prices, particularly at the entry-level end of the market, driven by a dramatic spike in memory chip costs. “This is just a taste of what’s to come,” warns analyst Anthony Scarsella. For the full year, IDC is forecasting a historic 12.9% collapse, which would bring the market down to its lowest level since 2013.

The downturn is hitting China’s three biggest brands hardest. Xiaomi shipments plunged 19%, Oppo’s dropped 10%, and Vivo’s fell 7%, reflecting their heavy exposure to emerging markets. Prices there have jumped by 40% to 50%, IDC notes, and some devices have been pulled the shelves altogether. Apple and Samsung, by contrast, posted modest gains, buoyed by strong demand for the iPhone 17 and Galaxy S26. The Korean group remains more vulnerable than its US rival, given its product lineup that stretches all the way down to the budget tier.

AI reshapes the chip market

Like other industries, smartphones are being squeezed by the rapid expansion of infrastructure for generative AI. This surge in computing power requires not just GPUs, but also vast amounts of memory. To meet booming demand, the industry’s three giants — SK Hynix, Samsung, and Micron — have shifted part of their production toward more profitable AI-oriented components, at the expense of DRAM and NAND memory used in smartphones, PCs, and gaming consoles.

According to TrendForce, 70% of memory production is now dedicated to AI. As a result, prices for all other components are rising. DRAM prices nearly doubled in the first quarter and are expected to rise another 60% in the second. Manufacturers do plan to expand capacity, but new production lines will not come online before 2027 or 2028 — and even then, there is no guarantee they won’t be primarily, or even entirely, allocated to AI memory.

A “structural reset” ahead

“While memory prices are projected to stabilize by mid-2027, they are unlikely to return to previous level,” anticipates Nabila Popal, an analyst at IDC. The market may be heading toward a “structural reset”, that could see sub-$100 smartphones disappear entirely. That would wipe out roughly 170 million handsets sold per year. Other segments will not be spared. Memory chips typically represent 15% to 20% of component costs in mid-range devices, and 10% to 15% in premium models.

For now, sales of these phones have held up thanks to inventory buffers and previously locked-in fixed-price contracts. But that window is closing. Samsung, for instance, has already raised prices on several models. If the trend persists, manufacturers will face tough trade-offs between passing costs onto consumers and protecting margins. Some may also scale back technical specifications. “They will have to reinvent themselves,” concludes Nabila Popal.

Photo: Apple

Barely launched in the advertising market, OpenAI is already setting its sights high. Buoyed by what it calls “encouraging early results,” the creator of ChatGPT is forecasting explosive, unprecedented growth, according to documents shown to investors and reviewed by U.S. outlets The Information and Axios. Within just five years, the company believes it can reach $100 billion in ad revenue — a level that would place it among the global top four, behind only Google, Meta and Amazon, but ahead of YouTube and TikTok.

OpenAI’s optimism rests on two pillars. First, the ever-expanding adoption of ChatGPT. Three and a half years after its launch, the chatbot already attracts more than 900 million weekly active users. The company projects that figure will reach 2.75 billion by 2030 — on par with Facebook or Instagram, and ahead of TikTok. Second, it is banking on a strategic position closer to users’ purchase intent, which could translate into higher conversion rates. Together, these advantages are meant to make the platform a must-have for advertisers.

$8 million in a month

For now, however, OpenAI is still in the early stages of its advertising push. A first pilot phase was launched in early February, initially limited to the United States. It has since been extended through at least the end of April and expanded to three additional markets: Canada, Australia and New Zealand. “We hope to continue expanding into many more markets this year,” the company says. Ultimately, ads are expected to be rolled out to all free users, as well as subscribers to its new low-cost Go plan.

In the first six weeks of testing, more than 600 advertisers ran campaigns on the chatbot. Revenue exceeded a $100 million annualized run rate—equivalent to just over $8 million in a single month. OpenAI, led by Sam Altman, is anticipating a dramatic ramp-up. It is targeting $2.6 billion in ad revenue this year, rising to $11 billion in 2027, $25 billion in 2028, $53 billion in 2029, and surpassing $100 billion the following year.

An advertising infrastructure to build

To scale, OpenAI still needs to build a full-fledged advertising infrastructure, which remains extremely limited today. The company has only just begun testing a self-service platform that would allow advertisers to create and manage their own campaigns, an essential tool to reach the millions of SMEs that underpin the ad models of Google and Meta. It has also lowered its minimum spend requirements and will need to strengthen its measurement tools to help advertisers assess campaign effectiveness.

That step will be critical to justify its pricing model. Unlike Google or Amazon, OpenAI has not opted for a cost-per-click model, shifting more risk onto brands. Its prices are also relatively high, exceeding those of streaming platforms, for example. Another challenge will be improving ad relevance. According to Adweek, some advertisers have reported click-through rates below 1%. A particularly complex task as the startup simultaneously seeks to expand ad coverage across a growing number of queries.

Ambitious or unrealistic?

Such a growth trajectory would be unprecedented. Google took 15 years to reach 12-digit ad revenues; Meta took 14. Amazon is not expected to get there until 2028, according to eMarketer, 12 years after introducing ads. But the context has changed. The online advertising market is far more mature, and the rise of AI is happening at a much faster pace. For now, OpenAI is also the only major player fully committing to this path.

Whether merely ambitious or outright unrealistic, these targets should above all be seen in the context of a potential IPO, likely by the end of the year. Beyond the headline figures, the goal is to convince investors that the company has a credible roadmap to monetize its vast base of free users. The stakes are high: achieving profitability will depend on it, especially as this audience generates significant costs due to the compute required for text generation.

Photo: OpenAI

For the past twenty years, Google has dominated the online advertising market after displacing Yahoo. But that supremacy could come to an end as early as this year: the search giant is on the verge of being overtaken by Meta, according to forecasts from research firm eMarketer. Based on its projections, the parent company of Facebook, Instagram, and WhatsApp could generate $243.5 billion in ad revenue, compared with $239.5 billion for its main rival. A marginal gap for now, but one expected to widen in 2027.

It is important to note that eMarketer’s estimates are based on net revenue. Roughly $80 billion is deducted from Google’s figures, notably excluding traffic acquisition costs (TAC). These expenses correspond to the sums the Mountain View group pays to partner sites that display ads sold via its platform. They also include a range of commercial agreements, such as payments to remain the default search engine on Apple’s Safari browser. By contrast, no such deductions are applied to Meta’s advertising revenue.

Two successive disruptions

Back in 2006, the handover between Yahoo and Google marked a major shift in a market then largely structured around traditional banner ads. By introducing sponsored links tied to user queries, the search giant imposed a model based on user intent, dramatically improving ad effectiveness. The goal was no longer to maximize audience reach, but to drive qualified traffic to websites. Campaigns were billed per click rather than per impression.

The potential transition in favor of Meta reflects the culmination of two successive disruptions. The first came in the early 2010s with the rise of social networks—a shift Google never fully managed to capitalize on, despite multiple attempts, including Google+. The second, more recent shift stems from the explosion in consumption of vertical video on social platforms, first driven by Snapchat’s ephemeral content and then by TikTok’s endless stream of short clips. Meta has been quick to integrate both trends across its apps.

The success of Reels

These shifts are reflected in the revenue growth trajectories of the two giants, which together are expected to capture more than 53% of the global market. Google’s growth remains relatively stable at around 12%. Meta’s, however, has sharply rebounded following a modest, but historic, decline in 2022. Its revenue is expected to grow by 24% this year, after 22% last year — a pace comparable to TikTok, despite starting from a much larger base. “This is unprecedented for a company of this size,” eMarketer notes.

The strong rebound of the company led by Mark Zuckerberg is largely driven by the success of Reels, Meta’s TikTok-inspired short video format. In October, Reels surpassed $50 billion in annualized revenue, becoming more lucrative than YouTube. Behind the scenes, the company is also benefiting from improvements to its advertising platform, particularly in campaign optimization tools and AI-generated creative. “Advertisers are seeing better returns on investment, which is attracting more ad budgets,” eMarketer adds.

Photo: Unsplash / Brett Jordan

Just two months after its unlikely merger with xAI, SpaceX is taking the next step. On Wednesday, the Elon Musk-led space company kicked off the process for an initial public offering. The deal, which could take place as early as June, may raise as much as $75 billion, primarily to fund the deployment of data centers in space, where generative AI models would be trained and run. That figure would set a new record, as would the targeted valuation of $1.75 trillion.

SpaceX has upended the launch market, notably with its reusable rockets. The company has also established its Starlink constellation, providing high-speed internet access worldwide. In just three years, its revenue has increased tenfold to $15 billion, and is expected to reach between $22 billion and $24 billion this year, according to Bloomberg. Reuters reports operating income of $8 billion. Growth prospects remain strong, driven by rapidly rising Starlink subscriptions and the anticipated opening of the mobile internet market.

Data centers “catapulted” from the moon

SpaceX is now also expected to serve Elon Musk’s broader ambitions in artificial intelligence, as reflected in its tie-up with xAI. The move appears designed to give the startup the financial firepower needed to compete with OpenAI, Google, and Anthropic. Known for its chatbot Grok, integrated into X, xAI is still generating limited revenue while pursuing heavy investment. It is posting steep losses — around $1 billion per month — that SpaceX’s rocket and satellite business will partly offset.

These new priorities are reflected in the company’s roadmap. Announced in February, the orbital data center project aims to bypass Earth-based energy constraints. “The only way to scale,” Elon Musk argues. To that end, SpaceX has asked U.S. regulators for permission to deploy up to one million satellites, eventually to be manufactured and “catapulted” from the Moon. Elon Musk also wants the company, in partnership with Tesla, to design and produce the GPUs that would be placed in orbit.

Selling a vision of the future

Ahead of the IPO, the merger with xAI cuts both ways. On one hand, it could help justify a record valuation by adding a high-growth business and attracting investors eager to bet on generative AI, especially if the listing comes before those of OpenAI and Anthropic. On the other, it risks weighing heavily on the company’s financials due to xAI’s massive losses.

Elon Musk’s core challenge will be to sell investors a vision of the future, much as he did with Tesla. On Wall Street, the carmaker’s stock is no longer closely tied to its commercial performance. Instead, it hinges largely on two projects: the Cybercab robotaxi and the Optimus humanoid robot. Similarly, SpaceX’s projected valuation cannot be justified by revenue alone. It will rest on the promise of AI, and by extension, orbital data centers and a space-based chip manufacturing ecosystem.

Photo: SpaceX

Rarely has a research paper moved markets this much. Over the past week, memory chip giants have seen their shares slide, wiping out part of the spectacular gains accumulated over the past year. The trigger: a new technology unveiled by researchers at Google, suggesting a significant reduction in the memory required to run generative AI models. The breakthrough could lower inference costs, while weighing on demand for certain chips.

The actual impact on memory manufacturers remains uncertain. First, because the technology is still at the research stage, with no guarantee of large-scale deployment or clear visibility on real-world benefits. Second, because efficiency gains could be offset by a surge in usage, driven by cheaper AI. Not to mention that demand for these chips still far exceeds supply. Market reactions nonetheless reflect a persistent tension between massive commercial upside and fears of a potential bubble.

Six times less memory

The rise of generative AI does not rely solely on massive computing power delivered by GPUs. It also depends on vast amounts of memory, provided by different types of chips — especially HBM (high-bandwidth memory), which is critical for training models. This voracious demand has triggered supply shortages, pushing manufacturers to reallocate capacity toward AI components and driving up prices. The ripple effects extend well beyond the sector, affecting chips used in smartphones and PCs.

In concrete terms, Google’s innovation targets the inference phase, the process of generating text or images. It takes the form of a model compression algorithm capable of drastically reducing memory requirements “with zero accuracy loss.” In tests conducted on several open-source models, memory usage was cut by a factor of six. “It is exceptionally efficient to implement and incurs negligible runtime overhead,” the researchers claim.

An overreaction?

These efficiency gains do not apply to the entire pipeline. Even if confirmed in real-world conditions, they would not translate into a sixfold reduction in overall memory needs. Still, Google’s compression method could significantly lower inference costs—a potentially crucial advance as the rise of AI agents is expected to multiply use cases. It could also enable models to run locally, without the expense of cloud infrastructure.

Analysts, however, believe market concerns about memory demand are overblown. They point to three main arguments. First, HBM chips, the most profitable segment, are unlikely to be affected. Second, efficiency gains could allow more models to run on the same amount of memory, rather than reducing total demand for equivalent usage. Finally, lower inference costs could further accelerate AI adoption, a dynamic known as the Jevons paradox.

Photo: SK Hynix