Exploding Token Costs: China Seizes the Opportunity

AI Costs Explode: Companies Ration Tokens and Search for ROI

In 2024, companies began their AI experimentation phase with open wallets and are now hitting a harsh reality: Some have already burned through their annual budgets in just three months. Uber blew its entire budget for agentic AI in March, Meta CTO Andrew Bosworth is issuing internal warnings against senseless token waste, and at a top financial institution, employees are burning through hundreds of thousands of dollars monthly for the simplest questions posed to premium models. Providers have now switched from all-you-can-eat models to usage-based billing, while Google is already processing 3.2 trillion tokens per month—seven times more than a year ago. What was intended as a signal to Wall Street is becoming a cost trap: companies are now rationing AI access, directing employees to cheaper models, and Microsoft is even limiting access to Anthropic Claude for developers, who are supposed to use internal tools instead. → Wall Street Journal

Synthszr Take: The party's over. After a year of 'tokenmaxxing'—employees using AI as an end in itself—the bill is coming due. One executive aptly compares it: 'If your daughter needs tutoring in algebra, you don't need to pay Albert Einstein.' This is the classic hype cycle, just at record speed: first, throw premium models at everything, then realize that 82% of the spending yields no measurable benefit. But the real story here is different: we are witnessing the birth of a new computer discipline. Companies are learning (the hard way) that AI integration is not a tech problem, but an organizational one. Those who build the right governance structures now and link token usage to business outcomes will have a massive lead in two years.

Xiaomi: Global AI Token Platform Cuts Prices by up to 99%

Xiaomi announces drastic price cuts for its MiMo-V2.5 API, while simultaneously ending its 100-trillion-token incentive program. Prices are dropping by up to 99%, while token usage for existing users is increasing 5- to 8-fold. The new pricing structure will be effective worldwide from May 27, 2026. The creator incentive program, which had been running since April, was terminated early after all 100 trillion tokens were fully distributed. The price reduction is technically enabled by optimized inference systems: with SWA-based HiCache, the data volume of the KV cache is reduced to one-seventh, while the number of storable tokens increases fivefold. Xiaomi is positioning the platform as a global alternative to Western AI providers with the goal of 'enabling more people to use better models'. → Hello China Tech

Synthszr Take: Xiaomi is making a classic China move here: pulverize prices, ramp up volume, take over the market. A 99% price cut with a simultaneous 5- to 8-fold performance increase—that's not optimization, that's a statement to OpenAI and Anthropic. The technical details (KV cache reduced to one-seventh) show real engineering work, not just pure subsidization. What Xiaomi calls 'MiMo Orbit' could be the beginning of a global token platform that undercuts Western providers on price. The early termination of the 100-trillion-token program suggests massive demand. The interesting question will be whether Xiaomi can maintain quality or if it's the old game: first flood the market, then raise the prices.

StepFun: 198-Billion-Parameter Model at a GPT-4o-Mini Price

StepFun has released a 198-billion-parameter Mixture-of-Experts vision-language model on Hugging Face that uses only 11 billion parameters per token. The model achieves 400 tokens per second with a 256k context window and offers three selectable reasoning levels. The kicker: at $0.20 per million input tokens, it's priced in the same range as GPT-4o-mini but delivers significantly more performance. It scores 79.2 on SimpleVQA (Search) and takes second place on SWE-Bench PRO with 56.3—the model can independently search multi-file repositories and generate working patches. Its availability on the StepFun Open Platform (global and in China), OpenRouter, and NVIDIA NIM makes it immediately usable for production. → AINews

Synthszr Take: It's a familiar story: world-class performance at a fraction of Western prices. 198 billion parameters sounds massive, but the sparse MoE architecture with only 11 billion active parameters is the real lever—it saves massively on compute for the same performance. The three reasoning levels (low, medium, high) are pragmatically designed: developers can switch between speed and depth depending on the use case. The fact that the model runs on Mac Studios with 128 GB of RAM shows the new reality: frontier AI is becoming a commodity. What makes me skeptical: the extremely low prices for this level of performance suggest massive subsidies. StepFun is likely burning capital for market share—the only question is how long that can last.

The Internet Rebuilds: For Machines, Not Humans

Cloud infrastructure has so far been geared towards humans: predictable clicks, scrolls, and streams. AI agents behave differently. They instantly launch hundreds of database queries, search documents, and call APIs—and then disappear just as quickly. Amazon is responding by rebuilding a core part of its cloud infrastructure. AWS has introduced a new generation of OpenSearch Serverless: a system specifically designed for the unpredictable work patterns of AI agents. It scales up in seconds when agents become active and scales down to 0 when they pause. Cloudflare reports that bots already account for 31% of all HTTP traffic. 'Non-human traffic will overtake human traffic in the first half of 2027,' predicts Cloudflare manager Lai Yi Ohlsen. → TechCrunch

Synthszr Take: The internet is becoming a machine infrastructure. AWS is now separating compute from storage—previously, at least one instance always had to be running, like a permanently rented parking spot. Now, customers only pay for actual usage. Google aims to become both the entry and exit point of the web by the end of the year (as I wrote after I/O). The consequence: anyone still building infrastructure for human click patterns today is developing for yesterday. Databricks and Snowflake are already positioning themselves as AI memory systems. This is a decision every executive can make tomorrow morning: infrastructure investments must be agent-first, otherwise you'll be paying for empty parking spots in 2027.

Anthropic and xAI: Who's Telling the Truth?

Anthropic has leased space on xAI's Colossus cluster—for $1.25 billion per month. This is the largest compute transaction in AI history. But while xAI's S-1-Filing clearly states a three-year term until May 2029, Elon Musk tweeted about a 180-day lease with a 90-day termination clause. The discrepancy between the SEC document ('the customer has agreed to pay a monthly fee through May 2029') and the CEO-tweet ('This is a 180-day lease') raises questions: either Musk misrepresented his own contract details, or the S-1-Filing is misleading. Both would be problematic during the Quiet Period before the Börsengang. Musk's justification: if compute becomes scarce, they want to keep the option to reclaim capacity. → Techpresso

Synthszr Take: $15 billion in annual revenue from a single customer—that's impressive even by Musk's standards. What appears to be confusion over contract details reveals the core problem of the AI economy: compute is the new currency, and whoever controls the GPU clusters dictates the terms. Anthropic is paying more for computing power than some DAX-listed companies make in revenue. The 90-day clause? A sword of Damocles hanging over Anthropic's model training. If xAI needs the capacity for its own projects, Anthropic is suddenly left without infrastructure. This is vertical integration through the back door, disguised as a cloud service. The SEC probably won't intervene (they rarely do with Musk), but the message is clear: in the AI era, compute providers are the new gatekeepers.

Opus 4.8: Prompting Strategies for Anthropic's Top Model

Linas Beliūnas presents a 31-page guide on how to get maximum performance out of Claude Opus 4.8. The key lever is the Effort-Level system with five tiers from 'low' to 'max,' where Opus 4.8 at minimum effort already achieves the peak performance of Opus 4.7 at maximum effort. The new Dynamic Workflows feature allows Claude to write its own orchestration scripts and spin up parallel sub-agents—Anthropic engineers have been using it internally for months. The cost remains at $5 per million input tokens and $25 per million output tokens. Beliūnas recommends: start with Sonnet 4.6, upgrade to Opus 4.8 for complex reasoning tasks, and use Haiku 4.5 for high-volume tasks. A practical tip: for xhigh or max effort, you should set a minimum of 64,000 tokens to give the model enough room to 'think'. → Linas from Linas's Newsletter

Synthszr Take: Anthropic made a statement with its $965 billion valuation on the same day as the Opus 4.8 release: the token economy is scaling up brutally. The Effort-Level system is essentially a compute discipline function—you pay for the thinking time you need. While everyone is talking about Agentic AI, Anthropic is delivering the necessary infrastructure with Dynamic Workflows: Claude writes its own orchestration scripts and manages sub-agents itself. This is reminiscent of the Jevons paradox: the more efficient the resource (in this case, reasoning power), the more it is consumed. At $25 per million output tokens and 64k token limits per run, we're quickly talking about double-digit dollar amounts for a single complex task. The math works out if the output generates corresponding business value—but the days of carefree prompt experimentation are over.

Linear: Product Development System for Teams and AI Agents

Linear is positioning itself as a 'new species' of product development tool, explicitly designed for collaboration between humans and AI agents. The system promises to integrate the entire workflow from idea to code review—with agents that can independently handle issues, write PRDs, and create pull requests. The demo shows a 'Codex' agent reacting in real-time to iOS performance issues and developing solutions on its own. Linear emphasizes three core promises: Purpose-built (oriented around the practices of world-class product teams), AI-native (agents as equal team members), and Speed-optimized (reducing overhead for higher velocity). Its customers include Ramp, GitHub, and OpenAI. → Techpresso

Synthszr Take: Linear is making the right noises: AI agents as full-fledged developer colleagues, seamless integration from PRD to PR, focus on speed. The demo shows a Codex agent analyzing and fixing an iOS performance issue in seconds—impressively staged. But the really interesting part lies elsewhere: Linear defines product development as a system where humans and agents use the same tools. Instead of selling AI as an add-on (like most), Linear is building the entire architecture to be agent-first. That could be the decisive difference: if agents truly become 10x more productive, they need an environment optimized for their way of working. Linear is betting that classic tools like Jira or Asana won't be able to handle this transformation. The bet could pay off—provided the agents actually deliver the promised productivity.

Slack: Automation to Combat Toolflation

79% of employees say their company does nothing to combat tool fatigue. Nearly one in five switches between apps over 100 times a day—costing more than 100 hours per year. Slack promises the solution with its Workflow Builder: automation right where work happens, with no coding skills required. New features like AI-generated workflows, conditional branching, and enhanced Salesforce integrations are set to make the platform the central nervous system for business processes. At Wayfair, automations are already saving 25,000 work hours annually, with a total of 3 million workflows running daily in Slack. → Techpresso

Synthszr Take: The problem is real—the solution is only partial. Combating tool fatigue with even more features in the master tool is reminiscent of the old IT wisdom: standards are great, everyone should have their own. Slack is evolving into the Microsoft Office suite of communication: in theory, it can do everything; in practice, teams use only 10% of its features. The real innovation lies elsewhere: 80% of Workflow Builder users are non-technical. This shows true democratization—when marketing managers can build their own approval processes without waiting for IT. The catch: each of these 3 million daily workflows is another dependency on Slack. Salesforce understands the game: first, monopolize communication, then build processes on top of it, and in the end, switching becomes prohibitively expensive.

The Orchestration Tax: Cognitive Bandwidth as the Bottleneck in the Agent Era

The cognitive overload caused by AI agents can be measured. Google engineers this week discussed a phenomenon they call the 'Orchestration Tax': the human developer becomes a serial bottleneck in a parallel system. Running 20 agents simultaneously doesn't mean 20 times the code gets produced. Every decision, every code review, every conflict resolution must go through exactly one processor: the human brain. The parallelism of the agents meets the serial nature of human attention. The result: developers feel more productive than ever and, at the same time, more exhausted than ever. → Nico Lumma from Five Things

Synthszr Take: The Orchestration Tax is Amdahl's Law in its purest form: the serial part (human judgment) severely limits the overall speed of the system. A developer with 8 agents doesn't have 8x productivity, but 8x context-switching costs. Each switch between agent outputs costs minutes, not microseconds like with CPUs. The solution isn't more discipline or longer workdays. It's in architecting one's own attention as a scarce resource. Anyone who ignores this will end up with superficial code reviews or what Addy Osmani calls 'cognitive surrender': accepting the agent's output because the mental energy for a proper evaluation is lacking. The ironic punchline: AI agents make us more productive at the part that was never the bottleneck.

Peak Hype: Buying a House with Anthropic or OpenAI Stock

A home seller in San Francisco is offering their three-million-dollar house in Duboce Triangle in exchange for shares of Anthropic or OpenAI. The offer has been live for 24 hours, and realtor Rachel Swann is being flooded with inquiries. The seller is a local luxury developer who believes in the two AI companies. The 232-square-meter property comes with 3-meter ceilings, custom built-ins, and remote-controlled solar skylights. The ad's lead line: 'Anthropic or OpenAI stock will be considered as payments.' San Francisco's real estate market and AI stocks have one thing in common: high demand, low supply, and exploding prices. → Business Insider

Synthszr Take: This is the perfect caricature of the current AI hype. A real estate developer is trading concrete for paper money from companies that have never paid a cent in dividends. The real punchline is elsewhere: both stocks are as illiquid as the house purchase itself—Anthropic and OpenAI only trade on secondary markets at astronomical prices. The seller is betting that the next funding round will double the valuation again (OpenAI is aiming for $150 billion). What passes for innovation here is actually a barter deal between two overheated asset classes. If this becomes the new normal in San Francisco, we should prepare for more creative payment methods—perhaps Tesla options for the next condo?