Money Circuit — Week of June 26, 2026
ABU DHABI BACKED FIRM MGX RAISES $50 BILLION TO BECOME THE PLANET’S BIGGEST AI INVESTOR AND IT ALREADY OWNS A PIECE OF ALL THREE TOP AI LABS
So while everyone was busy watching tech stocks crater this week, the guys in Abu Dhabi were quietly closing one of the largest investment vehicles in the history of capital markets. MGX, the UAE-based firm backed by Abu Dhabi sovereign wealth, confirmed they raised close to $50 billion from regional sovereign funds, global pension funds, and institutional investors to back artificial intelligence. Their goal is to hit $100 billion in total assets under management. They plan to deploy up to $10 billion a year in AI deals going forward.
But here is what makes this genuinely interesting rather than just a large number. MGX is, as far as anyone can tell, the only investor on earth that has placed bets into all three of the major AI labs simultaneously. OpenAI. xAI. Anthropic. They are in all three. That is not a hedge, that is a statement. It says we do not know who wins the AI race, nobody does, and we are buying all three horses at the same time.
Now stop and think about the scale of what is happening here. This is not a venture fund deploying $50 million into promising startups. This is a sovereign-adjacent vehicle dropping $50 billion into a single technological domain. There has never been anything quite like this in the history of technology investing. Not during the internet boom. Not during the mobile wave. Not during the semiconductor gold rush of the 1990s. This is genuinely new territory.
And the people doing it are not San Francisco tech bros or New York hedge fund managers. They are from the Gulf. They are capital pools that spent decades watching the oil economy, pricing long cycle commodities, and thinking in generational terms rather than quarterly returns. And they have made a calculation that AI infrastructure is the oil of the next 30 years. Whether you agree with that framing or not, you cannot argue with the conviction.
There is also a geopolitical dimension here that most coverage tends to skim past. The UAE is positioning itself as the neutral ground in the AI race, the Switzerland of artificial intelligence. By backing all the major American AI labs simultaneously, they are threading a needle that no single American investor could thread without creating conflicts. American VCs have to pick sides eventually because competition for talent, data, and compute is real. The UAE does not have to pick sides. They can be friends with OpenAI on Monday, xAI on Tuesday, and Anthropic on Wednesday, and nobody can complain because their checks cleared at all three.
The $10 billion annual deployment target also matters in ways that take a minute to sink in. At that pace, MGX will be among the most active AI investors in the world by volume within three years. That means board seats, information rights, and genuine influence over companies building the most consequential technology in human history. That is not a footnote. That is a major geopolitical fact that most people are not treating as seriously as they probably should.
The fund is financially closed, meaning the money is in the account and they have already started deploying portions of it. We will see where this goes. But if you were wondering whether the rest of the world was going to politely watch Silicon Valley own the AI future, Abu Dhabi just filed a very clear answer.
OPENAI AND BROADCOM BUILT AN AI CHIP CALLED JALAPENO IN NINE MONTHS AND IT COSTS HALF AS MUCH TO RUN AS A STANDARD GPU AND NVIDIA SHOULD PROBABLY WORRY ABOUT THAT
Eight months ago, OpenAI and Broadcom announced they were going to build a custom AI chip together. This week, they showed it to the world. Its name is Jalapeno. It is a reticle-sized ASIC built specifically for inference, which is the part of AI where the model is actually running and talking to you rather than learning from training data. The headline number is the one that matters: Jalapeno delivers roughly 50 percent cost savings compared to running the same workloads on standard AI GPUs.
If that holds up in production, that is not an incremental improvement. That is a structural shift in the economics of running AI at scale. And scale in AI right now means hundreds of millions of daily users, billions of queries a month, and compute bills that make enterprise software costs look like a rounding error.
Here is why this matters beyond the benchmark numbers. OpenAI runs on borrowed hardware right now. They lease compute from Microsoft, which runs on NVIDIA chips. Every time ChatGPT answers a question, NVIDIA makes money. Every time Claude answers a question, NVIDIA makes money. Every time any major AI product does anything useful for anyone, NVIDIA is sitting in the supply chain collecting its toll. The GPU monopoly has been one of the most durable and profitable positions in the history of technology, and OpenAI just took a very serious swing at dismantling it.
What is remarkable about Jalapeno is the timeline. Nine months from initial design to manufacturing tape-out for a high-performance AI chip. The standard development cycle for serious semiconductor work is measured in years. Years. The fact that they compressed it to nine months is either the most impressive engineering achievement of the year or a clear signal that OpenAI used their own AI models extensively throughout the design process. According to the announcement, they did exactly that. The chip was partially designed by AI. We are at the point where the technology is building itself.
Broadcom is the right partner for this because they have been doing exactly this kind of custom silicon work for years. Google has been running Broadcom-designed TPUs for nearly a decade and it has saved them billions in compute costs. OpenAI is running the same playbook, just faster. The deployment target is end of 2026, expanding from there, which means the actual financial impact will not show up in the numbers for another year or two. But the signal to the market is immediate and it is loud.
When you pair this story with the Qualcomm-Modular deal from the same week, which we will get to shortly, a pattern becomes unmistakably clear. The whole AI industry is working simultaneously to reduce its dependence on NVIDIA from multiple directions at once. From the hardware side through custom chips, from the software side through CUDA alternatives. Everyone is building bypass roads around the same toll booth. That is not a coincidence. That is a coordinated industry response to a pricing situation that has become too expensive to ignore.
NVIDIA will survive this. Their technological lead is real and their software ecosystem is deeply entrenched. But the era where NVIDIA was the only viable option for serious AI work is clearly ending, and Jalapeno is one of the cleaner examples of what comes next.
Source: TechCrunch / CNBC
$2.7 TRILLION GONE IN ONE MONTH AND THE STOCK MARKET IS FINALLY ASKING THE QUESTION EVERYONE WAS POLITELY AVOIDING
Here is a number that should get your attention. The Magnificent Seven tech companies plus Broadcom and Oracle have collectively lost approximately $2.7 trillion in market capitalization during June alone. The NASDAQ fell 2.2 percent on June 23rd. South Korea’s KOSPI briefly dropped 10 percent in a single session on the same day. SK Hynix and Samsung got absolutely hammered. The AI trade, one of the most profitable positions of the last three years, started looking very different this week.
The question investors are suddenly asking out loud is simple and has always been right there waiting to be asked: when does the AI capital expenditure actually pay for itself?
The spending numbers are staggering even by the standards of the biggest corporate buildouts in history. Microsoft, Google, Amazon, and Meta combined are expected to spend approximately $725 billion on capital expenditure in 2026. That is up 77 percent from $410 billion last year. Some analysts put the real number closer to $800 or $900 billion when you count Oracle and others. By 2027, the projections suggest the number could cross $1 trillion in a single year.
And investors this week basically stood up, looked at that number, and said: we need some proof that works before we keep bidding up your stock.
It is not that the market suddenly thinks AI is fake. Most sophisticated investors believe the technology is real and the transformation is coming. The question is timing and returns. If you spend $725 billion building data centers and buying chips and you generate $400 billion back in AI-driven revenue in the same period, the math does not work yet. The bill has been accumulating for three years. This week the market decided to acknowledge it.
There were also specific triggers for this particular moment. Key researchers started leaving major labs. Noam Shazeer, a co-lead on Google’s Gemini project and one of the original authors of the transformer architecture paper that started all of this, left for OpenAI. John Jumper, the Nobel Prize-winning DeepMind researcher behind AlphaFold, went to Anthropic. Talent migrations at that level spook investors because they suggest internal dynamics that nobody outside those companies can fully see, and internal dynamics tend to matter a lot in this industry.
There is also the Shiller price-to-earnings ratio for the US market, which crossed 40 for the first time since the peak of the dot-com bubble. That is a historical reference point that makes professional risk managers very uncomfortable very fast. And South Korea getting caught in the crossfire is a reminder that this is a global trade. Korean chip companies are deeply embedded in the AI supply chain, and when AI sentiment shifts in New York it shows up in Seoul within hours.
None of this means the AI boom is over. The underlying demand for what these companies are building is real and is growing. But the phase where investors would fund any AI story at any valuation on any timeline without asking about returns seems to be giving way to something more sober. The companies actually generating revenue and cutting costs with AI will be fine. The ones riding hype without underlying economics are going to have a harder time than they expected. The bill came due this week. Not the final bill, but a pretty significant installment.
Source: Yahoo Finance / Washington Post
QUALCOMM JUST SPENT $4 BILLION ON A STARTUP THAT LETS YOU RUN AI WITHOUT NVIDIA AND THAT IS EXACTLY AS INTERESTING AS IT SOUNDS
Tuesday was a very busy day in the AI money world. While the market was selling off, OpenAI was revealing a custom chip, and Abu Dhabi was closing a $50 billion fund, Qualcomm quietly announced it was buying an AI software startup called Modular for approximately $4 billion in an all-stock deal. And while this story did not get the headlines of the others, it might be the most strategically interesting one of the week.
Modular is not a household name outside of developer communities. But what they built is worth understanding. They created a programming language called Mojo, designed to make AI development faster and more efficient than what is currently available. They also built a platform called MAX that can optimize AI workloads to run efficiently across different types of hardware, including NVIDIA GPUs, AMD GPUs, Intel chips, Qualcomm chips, and whatever else you happen to be running. The whole idea behind Modular is that you should not have to be locked into NVIDIA’s ecosystem to run AI well.
This gets at what people in the industry call the CUDA problem. CUDA is NVIDIA’s software platform for programming AI. It works extremely well. It works exclusively on NVIDIA hardware. Developers have spent years learning it. Companies have built entire infrastructures around it. Switching away from CUDA is painful and expensive, which is a significant part of why NVIDIA has been able to maintain its pricing power even as competitors have tried to build alternative hardware. The software lock-in is as real as the hardware lead.
Modular was built specifically to give developers a credible path out of that lock-in. And Qualcomm, which makes chips and has serious ambitions in the AI inference market, just paid $4 billion to own that path.
The strategic logic is clean. Qualcomm has good hardware for AI inference at the edge and increasingly in the data center. What Qualcomm has not had is a compelling software story. When a data center customer looks at switching from NVIDIA chips to Qualcomm chips, the first question they ask is not about performance per watt. It is about whether they will have to rewrite everything. With Modular, Qualcomm now has a credible answer to that question. The MAX platform can help you migrate. The Mojo language gives developers a modern alternative path. That is a real sales tool that did not exist before Tuesday.
Whether $4 billion is the right number for what Modular had built is genuinely hard to say. They were not a company with obvious revenue metrics that would justify that price on a traditional financial basis. This is a strategic bet on where the AI software stack is heading and who will control the tooling layer that sits between hardware and applications. That layer has historically been worth enormous amounts of money to whoever owns it. NVIDIA understood this when they built CUDA. Qualcomm is betting that Modular can build the alternative version of that same layer.
In the same week that OpenAI showed Jalapeno, a hardware attack on NVIDIA’s position, Qualcomm executed a software attack on the same position. Two different companies, two different approaches, same target. If even one of these bets pays off significantly, the AI chip landscape looks very different in five years than it does today.
OPENAI FOLLOWED ANTHROPIC TO FILE FOR IPO AND NOW BOTH BIGGEST AI LABS ARE HEADING FOR THE PUBLIC MARKETS IN THE SAME YEAR AND THIS HAS NEVER HAPPENED BEFORE
If you were paying attention to the bigger financial picture over the last few weeks you might have caught both of these, but a lot of people missed one of them. Anthropic filed a confidential S-1 with the SEC on June 1st, targeting an October public listing at a $965 billion valuation. Then one week later, on June 8th, OpenAI followed. They also filed confidentially with the SEC. Both of the two biggest AI labs in the world, companies that between them arguably represent the center of gravity of the entire AI industry, are heading to the public markets in the same calendar year.
This has genuinely never happened before in technology. There has never been a moment where two companies of this level of significance, directly competing in the same core market, both went public in the same twelve-month window. You could try to find analogies in past tech cycles but they do not really hold. Google and Yahoo were not this close in competition. The internet companies that IPOd together in 1999 and 2000 were mostly not directly competing at the product level. OpenAI and Anthropic are racing for the same enterprise contracts, the same consumer subscriptions, and the same foundational model dominance. And they are both about to ask Wall Street to price that race.
The financial numbers involved are almost comically large. OpenAI closed a $122 billion funding round in March at an $852 billion valuation, the largest private fundraise in history at the time. Then Anthropic raised $65 billion in late May at a $965 billion valuation, briefly making them the most valuable private technology company on earth. Both of these companies, carrying these valuations, are now in conversations with the SEC about becoming public companies. Goldman Sachs, JPMorgan, and Morgan Stanley are already circling. The underwriting fees alone will be staggering.
What this means for regular investors is significant. Within the next year, you will potentially be able to buy a share of the AI race directly on a stock exchange. You will not need to be a sovereign wealth fund or a pension giant or a Sand Hill Road insider to own a piece of OpenAI or Anthropic. You will just need a brokerage account and an opinion about which one wins.
The number that will define both IPOs is not the revenue figure or the user count. It is gross margin. Anthropic is reporting $47 billion in annualized revenue as of May 2026, up from $9 billion at the end of 2025. That growth rate is extraordinary by any measure. But compute costs for running large language models at scale are enormous, and the gross margins on AI inference are not traditional software margins. They are structurally heavier than what Wall Street typically rewards at software multiples. Before investors can sensibly value these companies, they need to see the actual profitability dynamics in a public filing.
There is also what these two IPOs will do to the broader AI private market. When OpenAI and Anthropic are public companies with trading multiples and quarterly earnings calls, the private market valuations that have driven AI funding to record levels will have to make contact with public market reality. Investors will have a reference price. The somewhat magical thinking that has characterized private AI valuations for the last two years will have to survive comparison with what public markets actually pay.
Given what happened to tech stocks in June, including the $2.7 trillion wipeout we covered earlier in this column, that moment of contact between private AI optimism and public market skepticism is going to be one of the most interesting financial events in years. Get your popcorn ready. And maybe your brokerage account too.
Source: TechCrunch / Fortune