DX Today AI Daily Brief - Friday, June 26, 2026

DX Today AI Daily Brief - Friday, June 26, 2026

Today's briefing tracks an AI industry running at full tilt. Meta is quietly building a standalone prediction-market app, codenamed Antwerp, that lets users bet on real-world events using Llama-generated questions. OpenAI and Broadcom unveiled Jalapeno, OpenAI's first custom inference chip, taken from design to tape-out in about nine months. The infrastructure race accelerates as Hyperscale Data signs a 1.2 billion dollar AI compute deal in Michigan, SpaceX lands a 6.3 billion dollar Colossus compute lease with open-source lab Reflection, and Nvidia announces a record 35 new AI supercomputers across Europe. On the model front, Google DeepMind's Gemini 2.5 Pro Deep Think resets the science and reasoning leaderboard, while Anthropic lands Nobel laureate and AlphaFold creator John Jumper from DeepMind. The funding wires stay hot: French health insurer Alan raises 480 million euros, commerce-software maker Redo raises 81 million dollars at a 1.25 billion dollar valuation, regulated-industry agent startup Trase raises 107 million dollars, long-running-agent lab Sail Research raises 80 million dollars, and AI-evaluation company Patronus AI raises 50 million dollars, underscoring a cycle where capital flows to reliability, safety, and oversight as much as to raw capability.

DX Today - Your daily AI news briefing, delivered every morning.

Send us fan mail

#AINews #DXToday #OpenAI #AIInfrastructure #AIFunding

Transcript

SPEAKER_05 0:00

It's Friday, June 26, 2026. You're listening to the DX Today AI Daily Brief. Today, Meta builds a standalone app that lets people bet on the future. OpenAI unveils its first custom silicon, and Europe lights up a record number of AI supercomputers. Let's get into it.

SPEAKER_00 0:22

We begin with Meta. Internal documents reviewed this week show the company is developing a standalone prediction market app, internally codenamed Antwerp and FB Forecast, where users guess the outcomes of real-world events. The app, expected to carry the public name Arena, would lean on Lama, Meta's large language model, to automatically generate questions from trending topics, then serve what the documents describe as personalized market recommendations to each user. It marks Meta's most direct move yet into the booming prediction market space, a category that has drawn fierce competition and equally fierce regulatory scrutiny. Whether Arena ships as a real money product or a play money experiment remains unclear. But the ambition is plain. Turn everyday news into something you can wager on.

SPEAKER_05 1:11

From social bets to silicon.

SPEAKER_02 1:14

OpenAI has unveiled its first custom chip. Working with Broadcom, the company introduced Jalapeno, an inference processor it calls its first intelligence processor, architected specifically for serving large language models rather than training them. OpenAI says the chip was taken from initial design to manufacturing tape out in roughly nine months, which it claims may be one of the fastest advanced semiconductor development cycles ever, helped along by using its own models to accelerate parts of the design. Broadcom handles the silicon implementation, with Celestica helping industrialize the platform. Initial deployment is targeted for the end of this year, part of a broader push toward gigawatt scale data centers and tighter control of OpenAI's full stack.

SPEAKER_04 2:21

The infrastructure race intensifies. The build-out funding, it keeps growing. Hyperscale data disclosed a new master services agreement with a California-based NeoCloud provider for 20 megawatts of critical AI compute capacity at its Michigan data center campus. The company values the initial deal at roughly $1.2 billion, with a pathway to expand the footprint to 52 megawatts, which would push the total value above $3 billion over its term. It's a striking number for a single regional campus and a sign of how aggressively smaller players are racing to lock in power and capacity as demand from AI model providers outstrips supply. For Michigan, it signals that the data center gold rush is spreading well beyond the traditional hubs of Northern Virginia and Texas. And another mega compute deal.

SPEAKER_01 3:12

SpaceX is becoming a serious player in AI compute. The company signed a computing power deal with Reflection, an open source AI startup founded by former Google Deepmind researchers, worth up to $6.3 billion if it runs through 2029. Reflection will pay around $150 million a month beginning July 1st for access to NVIDIA's top-tier GB300 chips at SpaceX's Colossus 2 data center near Memphis. It's the fourth major Colossus Compute Lease, following deals with Anthropic, Google, and Cursor, and it shows how SpaceX has quietly turned spare data center muscle into a commercial cloud business. Both sides can exit with 90 days' notice after the first three months.

SPEAKER_05 3:59

Now to Europe's supercomputing surge.

SPEAKER_03 4:01

Nvidia used the ISC High Performance Conference in Hamburg to announce a record 35 new AI supercomputers now in development across Europe. The company calls it Europe's largest one-year expansion of supercomputing, spanning national centers, Euro HPC AI factories, universities, and industrial research labs across 23 countries. Together, the systems are designed to deliver hundreds of exaflops of AI performance, equipping more than 3 million researchers with next-generation infrastructure. The push reflects a broader European drive for technological sovereignty, a determination not to depend entirely on American hyperscalers for the compute that underpins modern science. From climate modeling to drug discovery, these machines are meant to keep European research competitive in an era where access to AI horsepower increasingly defines who leads.

SPEAKER_05 4:59

The Frontier Model Race tightens.

SPEAKER_00 5:02

Google DeepMind has reset the reasoning leaderboard. Its Gemini 2.5 Pro, running a new Deep Think Reasoning mode, posted benchmark results that now lead the field on science and graduate level reasoning. The model scored roughly 82% on GPQA Diamond, a brutal test of graduate physics, chemistry, and biology, surpassing rival frontier models, and it topped the charts on MMLU Pro as well. The picture isn't a clean sweep. Competing models still lead on software engineering and long horizon agentic coding, so teams are increasingly choosing tools by task rather than crowning a single champion. With prediction markets betting on more launches before months' end, the final week of June is shaping up as one of the most compressed model evaluation periods the industry has seen.

SPEAKER_02 5:53

A major talent move at anthropic. One of the most prominent scientists in AI is changing teams. John Jumper, who shared the 2024 Nobel Prize in Chemistry for Alpha Fold, the system that predicts protein structures from their amino acid sequences, is leaving Google DeepMind after nearly nine years to join Anthropic. Jumper's work helped turn AI into a genuine engine of scientific discovery, and his move signals Anthropic's growing ambitions beyond chatbots and into the natural sciences. It also underscores how fiercely the leading labs are competing for a tiny pool of elite researchers with marquee names now moving between rivals at a pace that would have been unthinkable a few years ago. For Anthropic, landing a Nobel laureate is both a research coup and a powerful recruiting signal.

SPEAKER_04 6:56

Turning to the funding wires, European Health Tech just landed a big one. Alan, the French digital health insurer, raised 480 million euros in a Series G round backed by investors, including Process, Index Ventures, and Teachers Venture Growth. Alan blends health insurance with an AI-powered app that handles claims, guides care, and nudges members toward preventive health. And the new capital is meant to deepen that AI layer and expand across Europe. The raise stands out at a moment when investors are increasingly favoring AI companies, tackling regulated, high-stakes workflows over broad consumer plays. Healthcare, with its enormous costs and stubborn inefficiencies, has become one of the most closely watched proving grounds for whether AI can deliver measurable results, not just impressive demos, more capital flowing to AI.

SPEAKER_01 7:52

Commerce software is drawing big checks too. Redo, which builds returns and post-purchase technology for online retailers, raised $81 million in a Series B led by Smash Capital, lifting its valuation to about $1.25 billion. Rido's pitch is that returns, long treated as a costly afterthought, can become a profit center when AI handles the messy logistics, predicts fraud, and turns refunds into exchanges that keep revenue in-house. The round reflects a quieter but durable trend. Investors backing AI that bolts onto real commercial pain points with clear return on investment, rather than flashy general-purpose tools. For merchants squeezed by thin margins and rising shipping costs, smarter returns handling is the kind of unglamorous problem that AI can actually move the needle on.

SPEAKER_05 8:44

AI agents for regulated work.

SPEAKER_03 8:46

Investors are betting on agents that can handle serious responsibility. TRASE raised $107 million led by Arch Venture Partners and Red Cell Partners to build AI agents for highly regulated industries, starting with healthcare and government-adjacent workflows. The premise is that the biggest enterprise opportunity isn't chat. It's automation that can execute multi-step tasks reliably in environments where a mistake carries real legal or financial consequences. That demands agents with strong guardrails, auditability, and human oversight baked in from the start. The size of the round, unusually large for an early stage company, signals deep conviction that the next wave of AI value lies in trustworthy execution. It also reflects how much of today's capital is chasing reliability over raw capability.

SPEAKER_05 9:43

Two more startups to watch.

SPEAKER_00 9:45

The agent Gold Rush continues. Sale research raised about $80 million across the seed in Series A at a valuation near $450 million, with Kleiner Perkins leading the later round and Sequoia leading the seed. The company is building what it calls long-running AI agents, systems designed to work autonomously on complex tasks over hours or even days rather than answering a single prompt and stopping. It's one of the hardest problems in the field because errors compound the longer an agent operates without a human in the loop. The heavyweight investor lineup suggests Silicon Valley believes durable, long horizon autonomy is the next real frontier, and that whoever cracks reliability at that scale could unlock a vast new category of AI software.

SPEAKER_05 10:33

And finally, guarding the guardrails.

SPEAKER_02 10:35

As companies rush AI into production, the tools that watch over it are cashing in. Petronas AI raised $50 million in a Series B from Greenfield Partners to expand its platform for evaluating and monitoring AI systems. Patronus helps enterprises catch hallucinations, test model behavior, and verify that deployed AI actually does what it's supposed to, an increasingly urgent need as autonomous agents take on real decisions. The RAIS captures a defining theme of this news cycle. The money is flowing not just to the models, but to the safety, reliability, and oversight layers around them. It's a sign that the industry is maturing, shifting from a question of what AI can do toward whether we can trust what it does.