Open-Source v Closed Systems

The CEO of Cerebras, a chip company, was on the Oddlots podcast and noted the debate between whether the closed source (U.S.) or open source (China) ecosystem will become more prevalent in AI. He didn’t want to hazard who would win although he felt that the Chinese models were not too far behind in most use cases. They have clients that will fork an open-source model and start using it with their inference-specific chips.

China’s AI ecosystem has 600 million users and deployment efficiency with Stanford University saying the performance gap is “effectively closed”. 59% of Chinese PhD students in America now plan to return home, and there have been high profile departures from Google and OpenAI back to China. Government linked investors went from less than 10 AI deals before 2018 to over 140 in 2025. DeepSeek’s next-gen model matches or nearly matches OpenAI and Google on coding, agentic and knowledge benchmarks. Models from Moonshot, Xiaomi, and Zhipu shipped at similar capability levels in the last four months.

DeepSeek’s valuation surged to $50 billion raising a $7.4 billion round this week led by Chinese government investment funds. DeepSeek is hiring product managers and R&D engineers to build an AI coding agent “modelled on Claude Code” and “benchmark against Claude Code”.

The Chinese strategy is:

Phase 1: Match model capability despite chip constraints (DeepSeek V4 with Huawei chips) Phase 2: Undercut on price (DeepSeek pricing is 9x less expensive; token purchasers have taken Chinese models from 1% to 60% share on OpenRouter. OpenRouter is an AI infrastructure platform that gives developers access to hundreds of large language models through a single API; this allows developers to switch between models dynamically). Phase 3: Close the product and user experience gap to eliminate quality perception issues (in progress)

With companies using more AI tokens for coding, spend levels have mushroomed allowing Anthropic to post amazing revenue growth. CloudZero surveyed 500 software engineers and found that 45% are now spending over $100,000 per month on AI up from 20% in 2024. Enterprises are now deploying “advisor models”, cheap open-source models that handle bulk work and only call more expensive frontier models when it gets stuck. Figma is using open-source models to sell 20% to 30% token consumption reductions as a product feature. Google CEO, Sundar Pichai, said, “many companies are already blowing through their annual token budgets, and its only May.” Uber’s CTO said his team maxed out its AI budget in a few months in 2026. Microsoft’s Satya Nadella is forcing his engineers to use Co-Pilot and other coding tools rather than Claude Code due to costs. CTOs are making room in tech budgets for rising AI bills by squeezing SaaS budgets.

American labs are spending hundreds of billions on capex, training on expensive Nvidia chips and constrained by U.S. power grid capacity. Those costs are passed on to customers. Chinese labs facing export restrictions have been forced to use domestic alternatives and aggressively optimize for less compute. China also faces fewer power constraints as data centres powered by solar and hydro stand ready to take on more chips. This constraint became a strategy and now they are low cost.

Several industries such as banks, defence and regulated industries won’t touch Chinese models. But this is a narrow slice of the broader enterprise market. Most others are wondering about their token costs and starting to design around high frontier costs. The “advisor model” suggests customers are using cheap models as the default and treating the frontier as an expensive fallback. OpenAI has said that pricing is not a customer’s Top 10 concern, but recent behaviour indicates otherwise.

If DeepSeek matches frontier capability at one-ninth the cost, it will be interesting to see whether there is a margin collapse in enterprise AI such that the only defensible moat for the labs will be regulated industries, or whether institutional inertia gravitates companies into paying premium prices despite lower cost options. A good tell could be traditional SaaS companies. If AI-developed software leads to SaaS volume and pricing declines, it probably means enterprises are also optimizing on AI token costs. This could lead to margin compression for Anthropic and OpenAI down the road to preserve market share.

The K-Shaped Economy. AI stocks dominated the earnings season. Nvidia continues to grow its data centre revenue (77% growth year on year). Anthropic achieved massive revenue production in the last quarter and will become profitable years earlier than expected. Japan AI and memory stock momentum was trading at 5x-6x Sharpe ratios. Softbank surged limit up on AI news. Korean memory makers, Hynix and Samsung, have surged. Lenovo surged 20% in Hong Kong as its AI revenue surged 84% year on year. Cursor hit a $3 billion run rate in April up from $2 billion in February. AMD’s CEO said the CPU market would explode at a 35% CAGR for the next five years thanks to AI.

But the non-AI economy struggled. Walmart noted customer pressure saying, “The number of gallons that customers fill up with when they come to our fuel stations fell below 10 for the first time since 2022, that’s an indication of stress.” Kraft Heinz’s CEO said, ‘consumers are literally running out of money toward the end of the month.” Elf Beauty rolled back price increases as consumers witness macro pressure. H&M cut estimates below consensus with its gross margin bull case ending. The UK government is contemplating price freezes on important goods at grocery stores. Travel agency revenues showed sharp contractions in volume. Easyjet posted warnings about a slow-down in summer travel.

The AI boom may be masking a consumer recession.

Emerging Market Demand Destruction Kept Inflation in Check. I’m not sure what to make of the Iran deal that is almost done but not done. Perhaps watching the Lego shorts on X is the best way to track the negotiations. But while inflation has crept up, it hasn’t been as much as anticipated. This has made the market sanguine about the risks from closure in the Strait of Hormuz. But one thing to note is that in emerging markets there has been significant demand destruction of about 5m barrels oil/day. These countries did not have strategic reserves, and its citizens are more price sensitive to oil shocks. This has helped balance the market. But as strategic reserves deplete, it will soon impact Western economies too which may be why President Trump is trying to accelerate a deal.

If we get past the oil shock, two thematics are the industrial/credit cycle accelerating and easing expectations due to the weaker main street economy. The AI capex boom has been increasing inflation providing some tough decisions by the Fed in the months ahead.

Private Credit Defaults Accelerating. Private credit default rates hit a record 6% according to Fitch Ratings, with 81 first-time defaulters over the past year, the highest since tracking began, confirming the systematic stress we've been monitoring across the industry. Most concerning is that over half of these defaults involved borrowers switching to payment-in-kind (PIK) arrangements, meaning they're paying interest with more debt rather than cash, essentially kicking the can down the road while their debt loads grow larger. Consumer products companies saw defaults spike to 11.1% from 5.9% a year ago, while healthcare led in first-time defaults, and seven of April's ten defaults involved maturity extensions under stress all indicating that borrowers are struggling to meet their obligations as the $330 billion software debt maturity wall approaches and AI disruption threatens portfolio companies. This 6% default rate, combined with widespread PIK conversions masking even deeper distress, validates warnings from industry leaders like Jamie Dimon and Lloyd Blankfein that private credit's relaxed underwriting standards during the boom years are now producing the "higher than expected losses" and systematic problems that could ripple through the $300+ billion in bank exposures and trillion-dollar insurance company holdings backing American retirement security.

SpaceX News. Some interesting data-points for SpaceX’s upcoming $1 trillion or $2 trillion IPO. SpaceX is projected to generate $22 to $30 billion of revenue in 2026. It says its enterprise AI market opportunity is $26.5 trillion with no timetable for when this revenue materializes. Gartner thinks this market is $3.3 trillion by 2027. Citigroup says $4.2 trillion by 2030. The estimate is 6x-8x higher than the market research firms and represents a number close to U.S. GDP.

Historical performance data is troubling. Of 36 mega-IPOs ($15 billion+) only 9 outperformed the S&P. Companies with price to sales over 40x average negative 45% returns over the next three years. An average IPO lags the market by 25% after the initial pop.

Bitcoin Problems. Michael Saylor’s Strategy has bought more Bitcoin this year than the entire global mining network has produced. There has been a reduction in retail enthusiasm (they are now talking about semiconductors), hedge fund arbitrage and miner accumulation. But Saylor is there buying Bitcoin with dedication. This works if Bitcoin doesn’t fall far. But the more dominant he is in the market, the more fragile the thing he’s propping up. He may end up being Bitcoin’s biggest point of failure.

End Note

Jim Cramer came out and said, “It’s a new era. Semis are now in charge.” In 1999-2000, Cramer was a vocal proponent of “new economy” stocks and argued that traditional valuation metrics didn’t apply to Internet companies. He said this was the permanent nature of tech transformation. In March 2008, he said, “Bear Stearns is fine. Don’t move your money from Bear.”

Eric Schmidt gave a commencement speech at the University of Arizona suggesting graduates help shape AI rather than fear it. He was repeatedly booed. Real estate developer Gloria Caulfield told University of Central Florida graduates, ‘The rise of artificial intelligence is the next industrial revolution” and that AI should be embraced like past disruptive technology. She was also booed. Scott Borchetta, known for helping Taylor Swift’s career, said AI is transforming production and that graduates should treat AI as a tool rather than fear it. Graduates also booed, with Borchetta saying, “deal with it” because it’s not going away. However, Steve Wozniak was a notable exception saying, “you have AI, actual intelligence” and saying how engineers figured out how to make a brain and it “takes nine months”. He got a lot of cheers.

Omar Sayed