Uber blew through its entire 2026 AI coding budget by April. Microsoft revoked its developers’ Claude Code licences six months after enabling them. One company reportedly ran up a $500 million Claude bill in a single month after forgetting to set usage limits. A Priceline employee told TechCrunch that a routine Cursor contract renewal came back four to five times more expensive.

The pattern is the same everywhere. Per-token prices have collapsed, but the push for autonomous AI agents has sent consumption through the roof. Companies that gorged themselves on all-you-can-eat subscriptions in early 2025 are now scrambling to understand where the money went, and whether any of it produced a return.

The paradox in numbers

GPT-4-equivalent performance now costs roughly $0.40 per million tokens, down from $20 per million in late 2022. That is a 98% reduction. Yet enterprise AI bills have risen by an estimated 320%, according to multiple industry analyses. The average enterprise AI budget has grown from $1.2 million per year in 2024 to $7 million in 2026.

The culprit is volume. Agentic AI tools released since November 2025, including Anthropic’s Claude Opus 4.5, OpenAI’s GPT-5.1, and Google’s Gemini 3 Pro, have multiplied token consumption per task. A simple linear workflow in 2023 cost about $0.04 per interaction. An orchestrated agentic system in 2026 costs roughly $1.20, about 30 times more. Individual engineers at Microsoft were reportedly spending between $500 and $2,000 a month on tokens before the licences were pulled.

Nicholas Arcolano, head of research at engineering management platform Jellyfish, told TechCrunch that per-developer consumption has risen roughly 18.6 times in nine months. Engineers who used the most tokens were about twice as productive as lighter users, but they spent 10 times the tokens to get there. “Whether extreme spend pays off comes down to the ultimate business value of shipped code, which most companies still can’t measure,” Arcolano said.

From tokenmaxxing to guardrails

“Six months ago, I would have a conversation with a customer and it would be all about ‘What can it do? Is it good enough?’” Alexander Embiricos, OpenAI’s head of enterprise, told TechCrunch. “Now the conversations are about, ‘We’re spending so much. What visibility do you have? What token controls do you have?’”

J.R. Storment, executive director of the FinOps Foundation, described the shift bluntly. “In April and May, I started hearing from companies: ‘Oh my god, we are 3x over our entire 2026 token budget and it’s only April.’ The whole conversation shifted from tokenmaxxing and ‘go fast’ to ‘we need guardrails, how do we control this?’”

Priceline’s senior director of IT finance, Chris Reed, drew a comparison to the telecom billing era. “It’s like the crack-cocaine epidemic. They let you try it to get you hooked, and now you’re kind of beholden to it.” The company has begun placing token limits on certain groups. Reed said he is already seeing discrepancies between vendor-reported usage and Priceline’s internal data.

The Tokenomics Foundation

It is against this backdrop that the Linux Foundation this week unveiled plans for the Tokenomics Foundation, a new standards body aiming to bring the same cost discipline to AI tokens that FinOps brought to cloud spending.

The Foundation plans to build a canonical definition of “tokenomics,” open standards for AI token usage and billing, and new metrics including cost-per-intelligence and tokens-per-watt. A formal launch is planned for July. Nishant Gupta, chief availability officer at Salesforce, said in a statement that “token economics is fundamentally more abstract and opaque than anything we’ve managed at this scale before.”

The challenge is enormous. “Tracking cloud costs is a hundreds-of-millions-of-rows-a-month data problem,” Storment said. “Tracking token costs is a trillions-of-rows-a-month data problem.”

A market forms around the problem

Startups and established vendors are racing to fill the gap. Pay-i tracks and optimises AI spending. Paid lets developers bill based on actual value rather than subscription fees. Jellyfish, Waydev, and Faros AI provide agent monitoring to prove the ROI of developer tools. Ramp has moved into AI spend management. Datadog and New Relic have added token-level observability.

Model routing is emerging as the primary cost lever. Factory, an enterprise AI coding startup, launched a model router this week that automatically picks the cheapest adequate model for each task. Vitaly Gordon, CEO of Faros AI, said frontier labs are already doing this internally. “The financial report for how much you spend on Anthropic, even if you call the Opus model, some of the spend will be on Sonnet or Haiku, because they are smart enough to do it,” he said.

Goldman Sachs projects global token usage will multiply 24 times by 2030. The companies already over budget need solutions now, and the Tokenomics Foundation’s first deliverable is still months away. As Gordon put it: “Maybe we created a steam engine, but we still haven’t figured out the assembly line.”

iFi Audio is back in the dongle DAC fight with the new GO link 2 Max, a compact USB-C DAC/headphone amplifier designed for smartphones, tablets, laptops, and PCs. Announced around High End Vienna 2026, the new model lands at $85 USD which puts it directly into one of the most crowded corners of personal audio.

And crowded is being polite.

The dongle DAC category is now packed with options from iFi, FiiO, Shanling, AudioQuest, Schiit Audio, Questyle, and enough other brands to make your phone’s USB-C port consider early retirement. AudioQuest has a new model coming as well, so clearly nobody got the memo that the boat was already full and starting to take on water.

Still, iFi has been at this long enough to know the assignment. The GO link 2 Max is not trying to be a desktop replacement, a battery-powered Bluetooth DAC, or a tiny slab of CNC-machined jewelry with a price tag that makes you clean your glasses, reload the page, and wonder if someone misplaced a decimal point. It is a wired USB-C dongle DAC with more output power, dual-DAC architecture, iFi’s S-Balanced technology, and app-based firmware support for under $100.

That might actually be a good deal if the sound quality has been improved and the cable can take the abuse.

Dual ESS Sabre DACs in a Tiny USB-C Package

The GO link 2 Max uses a dual ESS Sabre DAC architecture, with one DAC chip assigned to each audio channel. iFi says the design improves detail, definition, and instrument separation versus a single-DAC layout.

Format support is also strong for the price: PCM up to 32-bit/384kHz and native DSD256. That is more than enough for the overwhelming majority of users streaming from Qobuz, TIDAL, Apple Music, or a local hi-res library. Nobody needs to pretend they are casually commuting with 11.2MHz DSD files. Call your therapist if that’s actually something on your smartphone.

The GO link 2 Max also uses iFi’s GMT clock circuitry with a specialized crystal oscillator, along with ESS technologies such as Time Domain Jitter Eliminator. The goal is lower distortion, cleaner timing, and better clarity from a device small enough to disappear into a pocket.

More Power Than the Size Suggests

The headline number is up to 241mW of output power, which is a lot for something this small and affordable. That does not mean it will replace a proper desktop headphone amplifier, and nobody should expect it to drive planar headphones without some strain at higher levels.

But for IEMs, efficient dynamic headphones, and many portable over-ear models, 241mW gives the GO link 2 Max enough muscle to be more than a basic USB-C phone adapter with delusions of grandeur. Han Solo would understand.

In our review of the previous iFi GO link Max, the appeal was clear: it was small, solidly built, genuinely plug-and-play, and offered a lot more volume, resolution, clarity, bass texture, imaging, and separation than a basic laptop or phone headphone output.

It also brought dual ESS Sabre DACs, 32-bit/384kHz PCM, DSD256 support, and a 4.4mm balanced output to the sub-$100 category, which made the $79 price feel like someone at iFi had either lost a bet or found religion.

The limitations were also clear: the attached USB-C cable was a structural weak point, the 3.5mm output had less power than the 4.4mm jack, and high-impedance dynamic headphones were not always the best match.

The GO link 2 Max appears to stay focused on the same core idea, but with more output power, dual DACs, Dynamic Range Enhancement, THD compensation, and better software support through iFi Nexis.

That is the right direction.

S-Balanced Output, Not a 4.4mm Balanced Jack

One detail needs to be stated accurately: the GO link 2 Max does not appear to add a 4.4mm balanced headphone output. Instead, it uses iFi’s S-Balanced technology through its 3.5mm headphone output.

iFi says S-Balanced applies balanced circuit principles to a single-ended 3.5mm output to reduce channel crosstalk and improve separation. According to iFi, the implementation cuts crosstalk between channels in half.

That distinction matters because “balanced” gets thrown around in portable audio like free drink tickets at a trade show. This is not the same thing as a 4.4mm balanced output. It is iFi’s own approach to lowering noise and improving separation from a standard headphone jack.

For most users with 3.5mm headphones and IEMs, that is probably more useful than adding another cable standard to the drawer of shame.

Dynamic Range Enhancement and Lower Distortion

The GO link 2 Max also includes Dynamic Range Enhancement, or DRE, which iFi says adds up to 6dB of additional range between the quietest and loudest moments in the music.

iFi also claims its THD compensation reduces distortion by more than 50% compared to the original GO link Max. That is a useful claim, but again, the listening test matters. Measurements can tell part of the story. Headphones, IEM sensitivity, source device behavior, and volume control implementation will tell the rest.

iFi Nexis App Support, But Android Gets the Good Stuff

The GO link 2 Max supports the iFi Nexis app, which enables over-the-air firmware updates, selectable digital filters, and volume limiting.

There is a catch: iFi says those Nexis features are exclusive to Android devices. That means iPhone, iPad, and Mac users should not assume they are getting the same app-based control experience.

The two selectable digital filters are hybrid and linear, giving Android users some control over the DAC’s sonic behavior. Whether most listeners will hear a dramatic difference is another matter. Digital filters are useful, but they are not fairy dust. They tend to make subtle changes, not convert a $85 dongle into a $2,000 desktop DAC because someone tapped the right button.

Hardware Volume Control Is the Smart Move

One practical feature is the GO link 2 Max’s hardware-based volume control. iFi says this lets users adjust volume without reducing digital resolution in the way software volume control can.

That matters most with sensitive IEMs, where small volume changes and low noise are important. It is not the flashiest feature on the spec sheet, but it is the kind of detail that can make a portable DAC easier to live with every day.

Specifications Compared

The Bottom Line

The iFi GO link 2 Max is for listeners who want a real upgrade from a phone, tablet, or laptop headphone output without carrying a desktop DAC or another battery-powered box. For $85, it offers dual ESS Sabre DACs, up to 241mW of output, S-Balanced technology, hardware volume control, and hi-res PCM/DSD support in a tiny USB-C package.

The dongle DAC market is packed tighter than a CanJam elevator, but this one stands out by focusing on the basics: more power, cleaner conversion, and better control for IEMs and efficient headphones.

Where to buy: $85 at Crutchfield | iFi Audio

Summer Game Fest 2026 ended with a bang on Friday with the reveal of the final game of the Final Fantasy 7 Remake trilogy, Final Fantasy 7 Revelation. Not only did we get to finally see the trailer for the last entry, we were treated to gameplay footage as well. And we won’t have to wait long — it’s expected to come out next spring. 

The game picks up right after the previous game, Final Fantasy 7 Rebirth, and includes all the fan-favorite characters and ties up the loose ends of this retelling of the story from Final Fantasy 7. 

The trailer showcases the final chapter, in which Sephiroth wields incredible power and it’s up to Cloud and his allies to stop him. Featured in the video is the final character to join the team, Vincent Valentine, who appeared very late in the previous game, Final Fantasy 7 Rebirth. We also see the Weapons – Ruby, Emerald and Ultimate – that will be available in battle for those who want a challenge. Revelation also brings back the ability to swap characters on the fly during battles. 

The Summer Game Fest stream included a new addition to the game, Function Integrated Tactical Suitwear, or FITS. This system will unlock new movesets and battle boosts while allowing players to customize the look of their characters. 

The Final Fantasy 7 Remake series began in 2020 as a retelling of the original game from 1997.  The three games expand on the original’s story while modernizing the graphics and combat, resulting in a series that offers significantly more game to play.  

Final Fantasy 7 Revelation will launch in 2027 on PC, PS5, Xbox Series consoles and the Nintendo Switch 2. 

If you’re still using Office 2019 on your Mac, your time may be running out.

Microsoft has confirmed that from 13 July, Office 2019 for Mac will lose the ability to create, edit and save documents due to an expiring security certificate. While the apps themselves won’t suddenly disappear, they could become far less useful. This is especially problematic for anyone who still relies on Word, Excel or PowerPoint as part of their daily workflow.

The situation is unusual because Microsoft sold Office 2019 as a one-time purchase rather than a subscription. Many buyers picked it up specifically to avoid recurring Microsoft 365 fees and were happy to stick with a version that covered the basics without constantly adding new features.

Microsoft officially ended support for Office 2019 for Mac back in October 2023. Until now, that largely meant no new features or security updates, while the apps continued working as normal. The upcoming certificate expiry changes that.

It’s become an accepted truth amongst tapeheads that there’s no point looking at new hardware, because there’s only one tape mechanism being made anywhere in the world anymore, and that it sucks. [VWestlife] may enjoy German automobiles, based on the name, but he’s also a tapehead– and he took the time to demonstrate on YouTube that the accepted truth just ain’t so.

The supposed One Mechanism to Rule Them All in Lo-Fi is designed or made by Chinese company Tanishin. Certainly Tanishin does make a tape mechanism, but as [VWestlife] demonstrates with a few teardowns, there’s absolutely more than one on the market. That doesn’t mean any of the new offerings will out-compete your vintage Sony Walkman, but it does mean there are differences worth considering if you were to buy new.

Note that it is handhelds like the Walkman being talked about– it must be, since there are both slot-loading and flip-loading decks still being made, and even if you’re not a tapehead you should be able to tell that those won’t share the same part on the BOM.

With a few teardowns, he finds three separate mechanisms, followed by a deep-dive into the Tanishin. If you’re looking to buy a new walkman– or perhaps use its guts to build a mass storage device-– you might want to watch the whole thing to help you pick. On the other hand, the mechanism doesn’t matter that much, as he points out. It brings the tape over the head, but that’s not difficult. Everything else– from the motor that needs to draw the tape out evenly, to the pickup and the preamps and amplifiers–is where noise and poor quality sound tends to creep in, especially when something’s built to a budget.

Overall, [VWestlife] takes pains to point out that these ‘crappy’ new players aren’t any worse than the original Sony Walkman– we’ve just been spoiled by decades of better media than the humble compact cassette. That’s no slight against the cassette– people are still pushing its limits to this day, like this insanely fast vacuum-driven mechanism we featured.

Thanks to [Stephen Walters] for the tip!

For three years, Microsoft’s artificial intelligence story has been inseparable from OpenAI. The partnership — cemented by a cumulative investment exceeding $13 billion — gave Microsoft early access to the most advanced AI models on the planet, catapulting its Copilot products into the enterprise mainstream and adding hundreds of billions of dollars to its market capitalization. To the outside world, Microsoft’s AI strategy was OpenAI.

Mustafa Suleyman wants to change that narrative.

In an exclusive sit-down interview with VentureBeat at Microsoft Build 2026, the CEO of Microsoft AI disclosed that a contractual change with OpenAI roughly six months ago granted his division the formal authority to pursue what he openly calls “superintelligence” — using Microsoft’s own researchers, its own data pipelines, and its own custom silicon.

“We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence,” Suleyman said. “So this is very early days.”

The comment, delivered matter-of-factly backstage at the Fort Mason Center here, offers the clearest signal yet of a strategic inflection point unfolding inside the world’s most valuable public company. Microsoft is not abandoning OpenAI. But it is building something alongside it — and, eventually, something that could stand entirely on its own.

Microsoft’s first in-house model family signals a new level of AI ambition

The most tangible evidence of that shift arrived the same day. Microsoft announced a family of seven new AI models developed entirely in-house by its AI Superintelligence Team, spanning reasoning, code generation, image creation, transcription, and voice synthesis. The models — branded under the “MAI” family name — are Microsoft’s most ambitious first-party AI release to date.

The flagship, MAI-Thinking-1, is a 35-billion-active-parameter reasoning model that Microsoft says matches leading models in its weight class on key software engineering benchmarks and demonstrates advanced mathematical reasoning. Suleyman emphasized one point repeatedly: the model was trained from scratch on clean, commercially licensed data, without distillation from third-party frontier models — a direct, if unstated, contrast to the widespread industry practice of using outputs from competitors’ systems to train cheaper alternatives.

“We train our reasoning models from scratch,” Suleyman wrote in a blog post accompanying the announcement. “We don’t distill from other labs and we don’t rely on unlicensed or opaque data.”

The rest of the family fills out a multimodal portfolio designed for enterprise deployment: MAI-Code-1-Flash, a lightweight coding model built specifically for GitHub Copilot and VS Code; MAI-Image-2.5, which supports both text-to-image and image editing; MAI-Transcribe-1.5, which Microsoft claims is the most accurate transcription model available, operating across 43 languages; and MAI-Voice-2, a multilingual speech-generation system. All of the models ship through Microsoft Foundry, the company’s model-hosting and deployment infrastructure, and for the first time, developers can tune model weights themselves through third-party platforms including OpenRouter, Fireworks, and Baseten.

But Suleyman made clear in the interview that the seven models are a proof of concept, not a finished product. The real project is the lab itself.

“Our job is to make sure that when we look out to 2030 and beyond, we have the capacity not just to buy models from third parties, but to build the absolute frontier, the best models in the world,” he said. “That’s a long transition.”

What “set free” from OpenAI actually means for Microsoft’s AI future

To understand what Suleyman means by “set free,” you need to understand the unusual contractual architecture that has governed Microsoft’s AI efforts for years.

When Microsoft invested billions into OpenAI beginning in 2019, the partnership came with a specific arrangement: OpenAI would build the frontier models, and Microsoft would serve as the exclusive cloud provider, integrating those models into its products and reselling them through Azure. The deal gave Microsoft extraordinary commercial leverage — access to the world’s most advanced AI without having to build it — but it also created a dependency. Microsoft was explicitly barred from pursuing its own AGI research, and the agreement even capped how large a model the company could train, restricting it from building systems beyond a certain computing threshold measured in FLOPS.

That arrangement was formally renegotiated. As Fortune and Axios reported in November, a revised deal with OpenAI removed those restrictions, clearing the way for Suleyman to launch the MAI Superintelligence Team and pursue what he calls “humanist superintelligence.” The result, in Suleyman’s telling at the time, was a “best-of-both environment, where we’re free to pursue our own superintelligence and also work closely with them.”

By the time he sat down with VentureBeat at Build 2026, roughly six months had passed since that self-sufficiency effort formally began. Microsoft had already started shipping in-house models — including MAI-Image-2-Efficient, a lighter-weight image generation model released in April — but the seven MAI models announced at Build are the team’s most ambitious release yet: a full multimodal family spanning reasoning, code, image generation, transcription, and voice.

Even so, Suleyman does not view the shift as a rupture with OpenAI. He described Microsoft’s current position as one of abundance, not scarcity.

“There’s no immediate urgent need to fill a gap in three months’ time or six months’ time,” he said. “We have OpenAI, we have Anthropic, we have thousands of models inside Foundry. So there’s already a huge amount of optionality available to us.”

The framing is telling. Microsoft’s push into first-party frontier models is not born out of a crisis in the OpenAI relationship but out of a strategic calculation: as AI becomes the most consequential technology layer in enterprise computing, the company cannot afford to depend entirely on partners for the foundational capability. “Over the next five years, we have to be able to produce state-of-the-art frontier-scale models,” Suleyman said. “That’s our mission.”

Suleyman says the shift from chatbots to autonomous AI agents has already begun

If the seven MAI models represent the technical ambition, a new capability called Frontier Tuning represents the commercial logic. Announced alongside the models at Build, Frontier Tuning allows enterprise customers to customize MAI models using their own proprietary data, workflows, and domain terminology, all within their own secure compliance boundary. The system uses reinforcement learning environments — what Microsoft calls “training gyms for AI” — that let agents learn directly from real workplace tasks without affecting production systems.

The results Microsoft shared are striking. An MAI model tuned for Excel reportedly matches GPT 5.4 performance while operating at up to ten times greater efficiency. Early enterprise adopters are seeing similar gains: when tuned for one unnamed organization’s exacting standards, the MAI model achieved the highest win rate of any model tested at roughly one-tenth the cost.

Suleyman framed Frontier Tuning as part of a broader evolutionary stage — a move from intelligence to action. “We’ve basically moved beyond just conversation,” he told VentureBeat. “Now we’re moving to action.”

He introduced a new framework for thinking about that progression: the shift from IQ (factual intelligence) to EQ (emotional intelligence, or the ability to follow tone and style instructions) to what he calls AQ — the “Actions Quotient.” 

Future AI agents, in Suleyman’s telling, won’t just answer questions. They will log into enterprise software, navigate complex multi-application workflows, and execute tasks across Excel, Word, Teams, Jira, Adobe InDesign, and customer relationship management systems — just as a human employee would.

“You should be able to show up on day one and almost provision credentials to a new AI agent,” he said. “The model needs to be able to move across all of these different environments, and that’s actually the great strength of Microsoft.”

The Build 2026 announcements bore this out in concrete product terms. Microsoft Scout, the company’s first “Autopilot” agent, operates as an always-on background assistant built on the open-source OpenClaw technology. It runs with its own governed identity inside Microsoft Entra, so its actions are auditable and attributable. Windows 365 for Agents gives AI agents their own managed Cloud PCs, allowing them to interact directly with applications and browsers inside enterprise environments. And the Foundry platform received major updates — including hosted agents with sub-100-millisecond cold starts, a new Microsoft Agent Framework, and one-click publishing to Teams and Microsoft 365 Copilot.

Why Microsoft believes enterprise data is the next AI training frontier

Suleyman also articulated why he believes Microsoft’s position is uniquely defensible — and the argument has less to do with model architecture than with where work actually happens.

“We’ve sort of hoovered up all of the obvious pools of training data,” he said, referring to the industry’s early scramble to ingest the open web. “In the next phase, we actually want to be able to give these agents to companies to train on their specific tasks with the data that they have inside of their own big workflows.”

The claim is subtle but consequential. The first wave of generative AI was trained on publicly available text — books, websites, Reddit posts, code repositories. That data is now largely exhausted, and its use is increasingly contested in court.

The next wave, Suleyman argues, will be trained on enterprise-specific data: the internal workflows, decision traces, and institutional knowledge that define how real organizations operate. Microsoft, which serves 493 of the Fortune 500 through Azure according to Suleyman, is already embedded inside those workflows through Microsoft 365, Teams, Dynamics 365, and the broader Azure ecosystem. Frontier Tuning is the mechanism that converts that positional advantage into model performance.

“People underappreciate that that’s going to be the next domain,” Suleyman said.

The early partner list for Frontier Tuning reflects the ambition: Mayo Clinic, where Microsoft is co-creating a frontier AI model for healthcare using de-identified clinical data; EY, which is tuning a tax-advisory agent for deployment to 75,000 professionals globally; Land O’Lakes, where Frontier Tuning delivered what the company’s product development scientist called “meaningful improvements in grounded outputs and style compliance”; and Pearson, which is using tuned models to provide learning-science-aligned feedback in its Communication Coach product.

The Mayo Clinic partnership may be the most significant. Microsoft and Mayo Clinic are collaborating to build a healthcare-specific frontier model that combines Mayo’s clinical expertise and longitudinal patient insights with Microsoft’s AI capabilities. The model will be owned by Mayo Clinic and deployed first within Mayo’s own environment before being made available to other organizations through Foundry.

Microsoft’s custom AI chips and GPU buying spree reveal the scale of its compute advantage

None of this works without an industrial-scale compute infrastructure, and Suleyman was unusually candid about the hardware economics underlying Microsoft’s strategy.

“We are the largest buyer of GPUs on the planet,” he said. “We’re the largest buyer of GB200s and GB300s in the world.”

Microsoft will continue purchasing Nvidia accelerators “for many, many years to come,” Suleyman said. But the company is simultaneously building its own custom silicon. Maia 200, Microsoft’s second-generation AI accelerator, is already running in production across data centers in Iowa and Arizona, with deployments planned for Italy, Australia, and South Korea. According to Microsoft, Maia 200 delivers the best tokens-per-dollar-per-watt in the company’s fleet.

Suleyman put a finer point on the economics in the interview: Maia 200 is 30 percent more cost-efficient than Nvidia’s GB200, he said. And when Microsoft co-optimizes its own MAI models to run natively on Maia silicon, the company sees an additional 1.4x improvement in performance per watt. “It is going to be cheaper in years to come to build on MAI models with Maia 200 and Maia 300 inside of Azure,” he said.

That claim — if it holds at scale — has profound implications for the competitive landscape. It means Microsoft is not merely buying its way to AI dominance through Nvidia; it is building a vertically integrated stack in which its own models, running on its own chips, inside its own cloud, tuned on its customers’ own data, could offer performance and cost characteristics that no competitor can replicate.

Suleyman rejects the idea that AI models are becoming commodities

Suleyman also pushed back sharply against one of the most popular narratives in Silicon Valley: that AI models are rapidly commoditizing.

“A lot of people are saying models are commoditizing,” he said. “I don’t think that’s true.”

His argument hinges on what he calls “quality tokens” — the proposition that the composition, curation, licensing, and deduplication of training data matter at least as much as raw scale. Microsoft’s new MAI models, he said, were trained on a pre-training mix composed of approximately 50 percent high-quality code, with the remainder drawn from commercially licensed and carefully curated sources.

The result, he argued, is a distinct “lineage” of models optimized for coding, reasoning, and agentic behavior — fundamentally different from models optimized for consumer chat, cultural content, or multilingual breadth.

“We’re going to see very distinct lineages that reflect different training objectives of different companies,” he said. “Quality tokens matter more than just brute-force scale.”

This is a strategically important argument for Microsoft to make. If models are commodities — if any lab can match the frontier within months using cheaper compute and distilled training data — then the model layer becomes a race to the bottom, and Microsoft’s billions in compute investment offer no durable advantage. But if model quality is a function of data discipline, research depth, and institutional patience, then the lab-building approach Suleyman is pursuing becomes a genuine competitive moat.

He used a specific metaphor to describe that approach, one borrowed from optimization theory: the “hill-climbing machine.” The phrase describes a system that continuously improves — cycle after cycle — by applying more compute, better data, and sharper evaluation. “The goal here is to build what we think of as a hill-climbing machine,” he wrote in his blog post. “An organization that can continuously improve, cycle after cycle.” The metaphor is revealing because it describes a process, not a destination. Suleyman is not promising that Microsoft will build the world’s best model next quarter. He is arguing that Microsoft is building the system — the research culture, the data pipelines, the silicon co-optimization, the evaluation infrastructure — that will produce progressively better models over years.

Inside Microsoft’s five-year plan to become a self-sufficient AI superpower

The strategic picture that emerges from Suleyman’s comments — and from the full scope of the Build 2026 announcements — is of a company preparing for a future in which AI capability is not rented from a partner but generated internally, at scale, across every layer of the stack.

Microsoft still needs OpenAI. The partnership continues to power Copilot, Azure AI services, and ChatGPT’s infrastructure. Suleyman acknowledged as much, describing Microsoft’s portfolio of model providers as a source of strength, not a problem to be solved. 

But the direction of travel is unmistakable. With its own frontier models, its own custom silicon, its own reinforcement learning environments for enterprise tuning, and its own autonomous agent infrastructure, Microsoft is constructing a parallel path — one that, by 2030, could make the company a fully self-sufficient frontier AI lab embedded inside the world’s largest enterprise software platform.

“Our ultimate goal is what we call Humanist Superintelligence,” Suleyman wrote in his blog post. “That means advanced AI systems designed to serve people and organizations, not replace them.”

Whether that goal is achievable — or even clearly definable — remains one of the great open questions in technology. And Suleyman expressed more confidence than caution when asked about the trajectory of progress. “I really think we’re at the tip of the iceberg,” he said. “The models are so much more powerful than we know how to extract intelligence from them.”

But confidence and execution are different things. Building a frontier lab is not an announcement; it is a decade-long commitment that requires retaining elite researchers, maintaining scientific rigor under commercial pressure, and producing results that justify the staggering capital expenditure.

Google learned this with DeepMind — which Suleyman himself co-founded in 2010, before joining Microsoft — and even that lab, widely regarded as one of the best in the world, spent years navigating the tension between pure research and product delivery.

Suleyman seemed aware of the contradiction. “If you rush it, you’ll screw it up,” he said.

The sticker on his laptop reads: “Patience and urgency.” It is a paradox that Microsoft now has five years — and several hundred billion dollars — to resolve.

In Ithaca, Penelope’s road trip begins with a stop at a gas station, a call to a friend, and the discovery of a dude hog-tied and blindfolded in the trunk of her car. She seems surprised by the situation, but she closes the trunk, gets in the front seat and drives.

Ithaca is a road-trip RPG from The Pixel Hunt, the French studio behind the heartbreaking narrative game Bury Me, My Love, and it’s about more than just feeling the breeze through the open window and singing along to your favorite jams. The game’s protagonist, Penelope, is a 30-something environmental lawyer who’s fed up with a system that protects corporate polluters, and she’s joined the Earth Protection Association in an attempt to fight back in a new way. She’s on her way to a place called Ithaca to meet other members of the EPA when she discovers the boss of a huge petroleum company in her trunk. Apparently, she didn’t put him there.

The Pixel Hunt

The game plays out from the driver’s seat, where Penelope can look for clues about what’s she’s really doing, and also call and text her friends and family. Penelope has to pretend she isn’t getting swept up in a large-scale eco-terrorism plot while also navigating chats with the EPA organizers, building up interpersonal and deductive skills. There are more than 30 NPCs for Penelope to talk to, with a full cast of voice actors.

The landscapes that fill the windshield are procedurally generated scenes of wind farms, sprawling bridges, mountain passes and clean blue waterways, and the game ends in one of seven final destinations. Players are able to dictate Penelope’s path with their narrative choices, and ultimately decide whether she completes the EPA’s plans or backs out at the last minute.

The Pixel Hunt

“We’re not making Ithaca to give answers,” Pixel Hunt founder Florent Maurin said in the game’s Day of the Devs reveal video. “We’re making it because we’re living through questions that feel increasingly hard to avoid about power, about violence, about the world we live in and how far we’re ready to go to defend it. About what still feels possible when nothing seems to work anymore. And we hope Ithaca will be one way to sit with those questions for a while, carefully and comfortably from inside a car that keeps moving forward.”

Ithaca is tentatively scheduled to come out in 2027, and there’s a Kickstarter for the game live today, along with a Steam page.

Brian Chesky has spent years as an AI kingmaker. He met Sam Altman through Y Combinator in 2006, advised him on managing OpenAI’s hypergrowth, and helped broker Altman’s return to power after the board fired him in November 2023. He was reportedly considered for a seat on OpenAI’s board.

Now he is entering competition with his protégé’s company. Chesky plans to back a new AI lab of his own, Bloomberg first reported on Wednesday, with a focus on user interaction and design. He will remain as Airbnb’s CEO and will not lead the lab himself. The details are early-stage and could change.

Why Chesky is unsatisfied

The move reflects a frustration Chesky has voiced publicly for more than a year. He said last year that Airbnb had not struck an LLM partnership because existing products were not quite ready for what he wanted to build. His argument is that travel and commerce require a rich visual interface, not the text-based chatbots that OpenAI and Anthropic have popularised.

Airbnb has not been idle on AI. The company hired Ahmad Al-Dahle, who led generative AI work at Meta including the Llama model family, as chief technology officer in January 2026. It has rebuilt its app around a large language model for conversational search, automated 40% of customer support queries with an AI bot, and rolled out AI-generated listing details and review summaries. A voice-based assistant is planned for later this year.

But Chesky appears to have concluded that buying AI from frontier labs is not enough. He wants to build at the model layer, not just the application layer.

A growing pattern

Chesky is not alone. Brett Adcock launched Hark late last year with $100 million of his own money to build a universal AI interface, then raised a $700 million Series A at a $6 billion valuation. Hark is also emphasising user interaction and hardware, with the lead iPhone designer from Apple now heading its design effort.

Mira Murati’s Thinking Machines Lab is pursuing “interaction models” that process continuous streams of audio, text, and video in real time. The common thread is a belief that the frontier labs have focused on intelligence at the expense of interface, and that the next defensible layer sits between the model and the user.

The trend carries a broader implication. When founders of Chesky’s stature stop waiting for OpenAI, Anthropic, or Google to deliver what they need and start building their own research capacity, it signals that the application layer has run into the limits of what commodity models can provide.

The Altman dynamic

The personal dimension is hard to ignore. Chesky and Altman’s relationship spans nearly two decades. Chesky met Altman through Y Combinator, which incubated Airbnb. When OpenAI took off, Chesky began meeting regularly with Altman to advise on scaling a technology company. During the November 2023 board crisis, Chesky advised Altman on public relations and rallied support among Silicon Valley executives.

Now Chesky is building an operation that will compete, at least in part, with OpenAI’s own ambitions in user-facing AI. It is unclear whether the new lab will train its own models or build specialised systems on top of existing ones. But the direction is clear: Chesky wants proprietary AI research, not an API subscription.

What we do not know

Nearly everything about the lab remains unspecified. There is no name, no announced team, no disclosed funding amount, and no timeline. Chesky’s commitment to remaining at Airbnb raises questions about how much of his attention the new venture will receive, and whoever leads it will inherit a founding chair whom TechCrunch described as “known as a micromanager.”

What is clear is the thesis. Chesky has watched the AI lab landscape from closer than almost anyone outside it, and he has decided that the interface problem, making AI useful in rich, visual, consumer-facing contexts, is important enough to warrant its own research operation. Whether a part-time founder can build one that matters is the open question.

What The Puff produces up to 1,500 puffs a day

Singapore’s hawker scene is not short of curry puff stalls. But when Lim Yuan Ming, 24, and his older brother Brandon, 29, decided to start one, they weren’t trying to out-tradition the traditionalists—they wanted to give the curry puff a modern twist.

Growing up around their parents’ Teochew porridge stall in Bedok, the brothers had a front-row seat to the realities of the hawker grind. That experience shaped their decision to build something simpler, more scalable.

With their friend Oh Chin Jie, 31—a trained cook—they pooled S$10,000 and launched What The Puff at Changi Village Hawker Centre in Dec 2024.

One and a half years later, they have three outlets, a central kitchen producing 1,500 puffs a day, and have made S$500,000 in revenue for 2025, all while Yuan Ming is still finishing his finance degree at SUSS.

We spoke with Yuan Ming to find out how they got here.

Hawker blood runs in the family

The Lim brothers are no strangers to the hawker trade. Their parents have been running Lao Er, a Teochew porridge stall at Block 216 Bedok Food Centre, for over a decade. Yuan Ming grew up helping out—cashiering on weekends, loading toppings on porridge, and absorbing the rhythms of the trade without realising it.

But Teochew porridge, with its 20-dish daily spread prepared from the early hours, wasn’t the model they wanted to follow.

It was their father who pointed them toward curry puffs, even suggesting they experiment with adding cheese to the curry filling—a slightly unusual combination, partly inspired by the cheese-topped curry ramen trend.

When we tasted it, we thought, oh my god, it actually works. Then we just went for it.

Lim Yuan Ming

Advertisement

Once they had the idea settled, they brought in Chin Jie, Brandon’s friend from National Service, who had gained culinary experience at salad bar chain The Daily Cut.

Chin Jie took the lead on recipe development, while Brandon focused on operations and Yuan Ming handled the backend work, including administration, payroll, and supplier relations.

Six months to get the puff right

What The Puff currently sells five flavours, all wrapped in a thick, crumbly, old-school curry puff crust: the Original Curry Puff and Sardine Puff at S$2 each, and the Cheesy Curry Puff, Black Pepper Chicken Puff, and Char Siew Chicken Puff at S$2.50 each.

Simple enough on paper—but getting there was anything but easy.

R&D began at home in Oct 2024, two months before the first stall opened, and continued well after.

The trio spent weeks eating curry puffs with friends and family as unofficial taste-testers, producing up to eight iterations a day. After perfecting the basics, the Cheesy Curry Puff alone went through about 10 variations before they were satisfied.

The pastry proved the hardest part to nail, as small changes in ingredients, temperature, moisture, or even wind could affect the result. It took nearly six months of iterating before the recipe was finalised.

“We just ate a lot of curry puffs—to the point we got a little sick of them,” Yuan Ming said with a laugh.

From one stall to three

Although the early days at their first outlet at Changi Village were slow, it taught them a lot about business.

Sales were modest enough that the three founders spent over a month making the puffs themselves daily before hiring their first employee, and none of them drew full salaries for the first six months.

Soon enough, their determination paid off, with returning customers spreading the word about their funky puff flavours.

By late Jan 2025, they reinvested S$15,000 from the first stall’s earnings to open a second outlet at Block 216 Bedok Food Centre—just two months after their launch, and in the same hawker centre where their parents run Lao Er.

In Jul 2025, they opened a third outlet at Punggol Coast Hawker Centre in Punggol Digital District, funded with S$20,000. It has since become their strongest-performing outlet.

One surprise for the team was that curry puffs aren’t just a breakfast item. When they trialled extended evening hours at Punggol, demand stayed strong well past their initial 3PM closing time, prompting them to adjust operating hours accordingly.

“Whatever can go wrong, will go wrong”

But running a business is not without its ups and downs.

In Mar 2025, Brandon and Chin Jie were called up for reservist at the same time—right when Yuan Ming was in the middle of his exams. He ended up running the business largely alone for two weeks, working 14-hour days. Along the way, he also had to contend with a delivery failure and a refrigerator breakdown that ruined 400 puffs.

Whatever can go wrong, will go wrong. Now I know to have a plan for it.

Lim Yuan Ming

The team also eventually decided to close their Changi Village outlet.

Despite being their founding location, its distance from central Singapore made logistics challenging, and foot traffic was affected by nearby roadworks. In Jan 2026, they closed the stall and opened a new outlet at Haig Road Market & Food Centre instead.

Financial setbacks have also been part of the journey for the team.

Beyond the hawker centres, they brought What The Puff to GastroBeats, a food and music festival, earlier in 2025. It gave them room to experiment with flavours, introducing festival-only specials like chilli crab, mentaiko tuna, and truffle mushroom.

However, the pop-up came at a cost as they took a financial hit of around S$8,000. Yuan Ming doesn’t see it as a failure, though.

“It gave us the chance to figure out if events are something we want to pursue,” he said. The experience also connected them with other entrepreneurs further along in their journey, offering a perspective they couldn’t have gotten from behind a hawker counter.

45,000 puffs a month & counting

Recently, What The Puff made its first CBD appearance, expanding beyond the heartlands in late May. Through a collaboration, the brand is now stocked at Hypha Provisions at Raffles Place MRT.

The most significant operational milestone, however, came in Dec 2025, when What The Puff set up a central kitchen: a 100 to 200 sqft space staffed by four to five people to scale production.

Most of the puffs are now prepared there and distributed to outlets for frying on-site, with the Punggol outlet remaining the only one handling end-to-end production.

This has increased capacity to around 1,500 puffs a day, translating to roughly 40,000 to 45,000 puffs a month across all outlets. At S$2 to S$2.50 per puff, it is not surprising that the brand hit S$500,000 in revenue in 2025.

However, the central kitchen has also introduced new cost pressures, with manpower emerging as the biggest expense. The team is now exploring automation for certain parts of production, though Yuan Ming is careful not to lose what makes the product distinct.

Currently, all puffs and fillings are still made from scratch.

The dough is beaten the day before, then portioned and wrapped from around 8AM until early afternoon. “The texture you get from hand-forming the edges is not easily replicated by machinery,” Yuan Ming said. “That matters to us.”

Wrapping speed, at least, has improved dramatically—from 30 to 45 seconds per puff in the early days to under 20 seconds now.

What The Puff also makes sure not much goes to waste. Unfried puffs are kept frozen until needed, so the team only fries what they expect to sell. For damaged or leftover stock, they’ve partnered with food redistribution platform Food Tool, where excess puffs are listed at lower prices for anyone who wants them.

10 stalls in five years

For now, the trio prioritises stability over speed. The team wants to tighten production consistency, refine the menu, and get their systems solid before pushing further.

They have their eye on the West, North, and South of Singapore, as online orders have already been coming in from those regions, and a fourth outlet is planned for the second half of 2026. Their five-year goal is 10 stalls islandwide.

Collaborations are also on the horizon, though Yuan Ming is waiting until production is stable enough to take them on properly.

For anyone thinking of getting into the hawker scene, his advice is to expect chaos, plan for it anyway, and don’t underestimate how taxing the environment is. 

“It’s hot, oily, and things will inevitably go wrong,” he said. “But the government gives a lot of subsidies and grants to help people come in. If you want to try it, try it now—before the pressure to play it safe gets too loud.”

• Find out more about What The Puff here.
• Read other articles about Singaporean businesses here.

Featured Image Credit: HungryGoWhere/ @springtomorrow via Instagram