1. Anthropic, ARR Approaches $45B, Surpassing OpenAI
• Core Source
“OpenAI’s Q1 revenue was $5.7 billion, $900 million more than Anthropic’s $4.8 billion.”
“Anthropic’s Annual Recurring Revenue (ARR) has already approached $45 billion, with growth continuing at a rapid pace. As a result, the market is increasingly viewing OpenAI’s competitive edge as diminishing compared to the past.”
“Anthropic is projecting Q2 revenue of $10.9 billion. Anthropic’s annual revenue is now on track to exceed its previous forecast of $18 billion, and may even reach OpenAI’s target of $30 billion.”
“OpenAI’s adjusted operating margin (excluding SBC) for Q1 was -122%. This contrasts with Anthropic, which is projecting an operating profit for Q2.”
• Expected Impact
While Q1 quarterly revenue still favors OpenAI ($5.7B) over Anthropic ($4.8B), on an Annual Recurring Revenue (ARR) basis, Anthropic has already surpassed OpenAI, approaching $45 billion compared to OpenAI’s approximately $25 billion. ARR measures subscription- and contract-based recurring revenue on an annualized basis, meaning Anthropic’s recent growth trajectory is significantly outpacing OpenAI’s.
Even more notable is the profitability gap. OpenAI is losing $1.22 for every dollar of revenue generated, recording an adjusted operating margin of -122%, while Anthropic is projecting an operating profit as soon as Q2. This suggests Anthropic is not merely closing the revenue gap, but pulling ahead in terms of cost structure as well.
Anthropic’s rapid growth is underpinned by expanding enterprise penetration, broad infrastructure partnerships with all three major cloud providers — Amazon, Google, and Microsoft — and the growing competitiveness of Claude models in coding and agentic AI. Notably, the computing cluster lease agreement with SpaceX (worth $1.25 billion per month) and ongoing discussions around Microsoft Maia chip supply signal an aggressive push to secure compute capacity, which could further accelerate growth once supply constraints ease.
For OpenAI, ChatGPT’s weekly active users (WAU) have stalled at approximately 905 million, falling short of the internal target of 1 billion and signaling a slowdown in user growth. The prevailing market view is that the generative AI landscape is rapidly shifting from OpenAI’s dominance toward a two-horse race with Anthropic.
2. OpenAI Pursues $1 Trillion IPO — Passive Fund Mag7 Selloff Risk Looms
• Core Source
“OpenAI is planning to confidentially file for an IPO with regulators as early as tomorrow, aiming to go public around September.”
“The expected fundraising amount is $60 billion, with an anticipated valuation of $1 trillion — surpassing Saudi Aramco’s $25.6 billion — making it the largest IPO in capital market history.”
“JPMorgan projects that passive funds may need to sell approximately $95 billion of existing holdings in the top 8 Wall Street tech stocks to accommodate the new weighting. The impact is expected to overlap significantly with Mag7 names.”
“Passive funds will be forced to mechanically reduce their weightings in existing blue-chip holdings (such as Mag7) to make room for this massive pure-play AI asset, triggering dispersion in fund flows.”
• Expected Impact
OpenAI’s IPO — with an anticipated valuation of $1 trillion and fundraising target of $60 billion — is on track to become the largest IPO in capital market history, making it far more than a simple corporate listing. It is widely seen as an event that could restructure the supply-demand dynamics of global capital markets.
The core risk is mechanical selling pressure from passive funds. When OpenAI is listed and included in major indices, index funds will be required to automatically reduce existing holdings to make room for the new entrant. JPMorgan estimates this could trigger approximately $95 billion in selling across the Mag7 and other top tech stocks.
The Nasdaq’s “fast-track index inclusion” rule adds another layer of risk. This mechanism allows a company to be included in the index as soon as 15 days after listing, meaning if SpaceX, Anthropic, and other major AI companies list in quick succession, the supply-demand shock could be highly concentrated in a short window of time.
Beyond the near-term mechanics, the IPO could catalyze a broader reframing of AI investment. The AI investment thesis to date has been centered on infrastructure — semiconductors, data centers, and cloud providers. OpenAI’s listing would open a direct path to investing in foundation model companies, potentially diverting capital flows away from infrastructure and toward the model and application layer. This represents a medium- to long-term valuation headwind for the semiconductor and data center sectors.
3. AMD Invests $10B+ in Taiwan Ecosystem + Helios Full-Scale Ramp in H2
• Core Source
“AMD has announced plans to invest more than $10 billion across the TSMC ecosystem. The purpose is to expand advanced packaging capabilities for next-generation AI infrastructure.”
“Venice EPYC production is ramping on TSMC’s 2nm process, and Helios, equipped with the MI450X GPU, is targeting multi-gigawatt-scale deployment starting in the second half of 2026.”
“AMD, investing more than $10 billion in the Taiwan ecosystem to scale up Helios AI rack production.”
“Microsoft has officially designated AMD as its second AI accelerator solution platform — deploying MI450 racks by year-end and adding the MI500 series in 2027.”
“AMD projects the CPU market will grow at more than 35% annually over the next five years.”
• Expected Impact
AMD CEO Lisa Su’s announcement at a Taipei forum on May 22 marks a strategic inflection point for AMD, moving beyond its identity as a chip designer to asserting control over the broader AI infrastructure supply chain.
The $10 billion investment centers on ramping production of Venice EPYC CPUs on TSMC’s 2nm process and the MI450X GPU, alongside deepening advanced packaging partnerships with Taiwan-based firms such as ASE and SPIL. The Helios AI server system will enter full-scale production through Taiwan ODM partners Wistron, Wiwynn, and Inventec starting in H2.
Demand is already secured. Microsoft has officially designated AMD as its second AI accelerator platform, with MI450 rack deployments planned for year-end and MI500 series adoption in 2027. Meta and xAI are also cited as key customers adopting Helios as core compute infrastructure.
The backdrop is a structural push by hyperscalers to diversify away from Nvidia dependency. The top five North American cloud providers are actively expanding AMD purchases to reduce concentration risk in AI compute. Lisa Su projects the CPU market will grow more than 35% annually over the next five years, providing strong long-term visibility. AMD’s moves signal that real cracks may be forming in Nvidia’s dominant position.
4. Rubin Rack BOM Analysis — GPU Share Shrinks as Memory Cost Surges 435%, PCB +233%, MLCC +182%
• Core Source
“Morgan Stanley’s analysis of NVIDIA’s next-generation Rubin AI rack estimates the system price at approximately $7.8 million, nearly double the GB300, though the key driver of value growth lies in peripheral hardware components rather than the GPU itself.”
“The biggest beneficiaries within the Rubin rack are PCB (+233%), MLCC (+182%), and ABF substrates (+82%).”
“In particular, memory costs have surged 435% compared to the previous GB300 generation, with memory’s share of the total rack BOM estimated to expand to approximately 26%. Meanwhile, the GPU’s relative share is declining, reflecting a shift in the AI infrastructure value chain away from GPU-centricity toward memory, substrates, packaging, and passive components.”
“ODM dollar value-add is projected to increase by 35–40%.”
• Expected Impact
Morgan Stanley’s analysis signals a fundamental reshaping of the AI server investment framework. The Rubin rack’s system price of approximately $7.8 million is nearly double the previous GB300 generation, but the beneficiaries of that value increase are not GPUs — they are the surrounding components.
The most dramatic shift is in memory. Memory costs have surged 435% versus the GB300, reaching approximately $2 million per rack, with memory now accounting for roughly 26% of total rack BOM cost. This is the combined result of exploding demand for HBM for Rubin GPUs and LPDDR5X for Vera CPUs. PCB is up +233%, MLCC +182%, and ABF substrates +82%, with passive components and substrates broadly capturing significant value.
This provides a strong investment rationale for expanding focus beyond GPU and HBM to the broader materials, components, and equipment (MCE) value chain. As system complexity rises, capabilities in assembly, thermal management, power design, and rack integration are becoming increasingly critical, and ODM dollar value-add is projected to grow 35–40%. The structural evolution of AI servers — from simple compute devices into ultra-high-density power, network, and memory integration systems — is fully reflected in these figures.
5. Microsoft Maia 200 — Discussions Underway to Supply Anthropic with Proprietary AI Chips
• Core Source
“Microsoft has pledged up to $5 billion in support to Anthropic, while Anthropic has committed to spending $30 billion on Azure.”
“Microsoft is positioning Maia not as a replacement for frontier training GPUs, but as a cheaper alternative to NVIDIA chips for certain inference workloads.”
“Microsoft is one of Anthropic’s largest customers. The scale of Claude model usage for powering Copilot is at minimum $500 million.”
“Anthropic hopes to coordinate with Microsoft to ensure its requirements are reflected in the design of the next generation of Maia chips.”
• Expected Impact
The significance of these discussions lies in the fact that Microsoft is, for the first time, attempting to supply its proprietary Maia 200 chip to an external AI company. Until now, Maia has been deployed solely within Microsoft’s internal infrastructure. A successful supply agreement with Anthropic would mark a new milestone in the external expansion of its in-house chip ecosystem.
Microsoft’s strategic motivation is clear. Google has built a successful ecosystem around TPUs, and Amazon has achieved meaningful traction with Trainium — yet Microsoft remains heavily dependent on Nvidia, unable to break free from that structure. By positioning Maia for inference workloads, Microsoft aims to reduce Copilot operating costs while creating a virtuous cycle: Anthropic’s real-world workload requirements feed directly into the design of next-generation Maia chips.
For Anthropic, adding Maia to its existing lineup of Amazon, Google, and Nvidia chips would make it the only AI company utilizing proprietary chips from all three major cloud providers. This represents a strategy of diversifying compute supply to mitigate bottleneck risk, while simultaneously deepening relationships with each cloud partner. In a relationship where Anthropic has already secured $5 billion in investment from Microsoft and committed $30 billion in Azure spend, extending the partnership to chip design collaboration represents a meaningful deepening of strategic alignment.
6. SpaceX Plans 10GW Solar Cell Factory in Bastrop, Texas
• Core Source
“SpaceX is pursuing construction of a 10GW-scale solar cell factory in Bastrop, Texas. The facility is designed across two floors, each producing 5GW.”
“Musk said at the World Economic Forum in January that he viewed solar power as a solution to the energy bottleneck constraining the build-out of artificial intelligence infrastructure. He announced an ambitious goal for both Tesla and SpaceX each to build 100 gigawatts per year of solar manufacturing capacity in the US over the next three years.”
“First Solar Inc. is the nation’s biggest solar maker with an annual domestic production capacity of about 14 gigawatts for its thin-film technology.”
“There are just three domestic silicon cell facilities active, the report showed.”
• Expected Impact
To understand this announcement, one must look back at the vision Musk laid out at Davos in January. He declared solar power the solution to the energy bottleneck constraining AI infrastructure buildout, and set an ambitious goal for Tesla and SpaceX each to build 100GW per year of solar manufacturing capacity in the US within three years — a combined 200GW that would dwarf current domestic levels. This 10GW factory is the first tangible step in executing that roadmap.
The scale is what makes this significant. First Solar, the largest solar manufacturer in the US, has annual domestic production capacity of approximately 14GW — meaning SpaceX’s single factory at 10GW alone would represent 71% of that. Furthermore, only three domestic silicon cell production facilities are currently active, with most panels assembled using cells imported from overseas. Combined with the US government’s steep tariffs on solar imports from China and Southeast Asia, structural demand for domestically produced, de-sinicized solar cells in the US is set to surge.
There is, however, a key risk. The Chinese government has blocked the export of high-end solar manufacturing equipment from Suzhou Maxwell Technologies to Tesla. Musk attempted to resolve this during the Trump administration’s visit to China, but the outcome remains unclear, creating the possibility of equipment procurement delays that could affect the factory completion timeline. Looking further ahead, the story connects to SpaceX’s ambition to power orbital AI data centers with solar energy — the S-1 filing explicitly references a goal of 100GW of orbital computing capacity and emphasizes the need for a de-sinicized supply chain.
Comment [0]