Sign in Subscribe
AI Intelligence

OpenAI Plans ChatGPT Superapp as China Cuts AI Prices 99%

Vijay Bhagwati

• 5 min read
Share

The industry is pivoting from chat to agents while Chinese labs cut API prices by 99%.

June 8, 2026 | Reading time: 8 minutes | Issue #183

OpenAI is planning a major overhaul of ChatGPT that would turn the chatbot into a "superapp" integrating coding tools and AI agents, the Financial Times reported on June 7. The redesign would embed software development, browser automation, and task-delegation agents inside the same interface, giving OpenAI a platform to sell higher-margin products atop a free chat tier. One employee told the FT that "chat is dead" as a standalone product. The move comes as OpenAI prepares for a potential IPO and needs to demonstrate revenue streams that do not depend on per-token API pricing. Bloomberg and TechCrunch confirmed the plans, noting that OpenAI has already rolled out a "Dreaming" memory upgrade and expanded Codex integrations in the past week. The risk is execution: superapps rarely succeed outside Asia, and OpenAI has little experience building integrated software suites. The reward is that the first western AI platform to own the application layer could capture margins that currently flow to Microsoft, Google, and independent agent startups.

Source: Financial Times

PhysicsX raises $300M Series C at $2.4B for industrial AI

PhysicsX raised a $300 million Series C at a $2.4 billion valuation, Bloomberg reported on June 8. The London-based startup uses AI to design physical parts for jet engines, semiconductors, and other industrial applications. Temasek led the round. PhysicsX does not build chatbots; it builds geometry. The bet is that generative models trained on physics simulations can replace months of manual CAD iteration with minutes of GPU inference. If the model generalizes across industries, the company sits on a wedge between incumbent engineering software and the hyperscalers' cloud AI platforms.

Source: Bloomberg

NVIDIA and SK Hynix sign multi-year next-gen memory pact

NVIDIA and South Korea's SK Hynix signed a multi-year agreement to develop next-generation memory chips for AI infrastructure, Bloomberg reported on June 7. The chips are tailored for NVIDIA's roadmap, including the Vera Rubin generation. SK Hynix already dominates high-bandwidth memory; the pact makes NVIDIA a co-designer rather than a buyer. The deal tightens NVIDIA's grip on the memory layer at a moment when Chinese startups are trying to build domestic alternatives.

Source: Bloomberg

Apple held secret meeting admitting it was behind in AI

Bloomberg reported on June 7 that Apple held a secret meeting in early 2025 where executives acknowledged the company was behind in AI. Tim Cook became personally involved in the roadmap, and WWDC 2026 is expected to unveil a major Siri reboot alongside AI feature reversals. The admission explains why Apple's AI releases to date have been cautious: the company knew it was trailing Google and OpenAI internally before consumers noticed. The question is whether a year of focused development is enough to close the gap on agentic assistants.

Source: Bloomberg

India and UAE sign Cerebras supercomputer deal to bypass US clouds

India and the United Arab Emirates signed an agreement on May 15 to deploy a Cerebras AI supercomputer in India, Rest of World reported on June 1. G42, backed by Abu Dhabi's Mubadala, will install 64 Cerebras systems, giving India a second path to AI compute that bypasses Amazon, Microsoft, and Google. Cameron Kerry, former acting US Commerce secretary, called it "India's pragmatic approach to AI sovereignty." Separately, TCS reported $2.3 billion in annualized AI services revenue in the first quarter of 2026, as Indian IT firms pivot from back-office outsourcing to AI deployment.

Source: Rest of World

Open-Source Pulse

DeepSeek's V4 Pro triggered a domestic price war so severe that Xiaomi cut MiMo-V2.5 API prices by as much as 99%, the South China Morning Post reported on June 8. MiniMax launched subscription tiers to retain users. The cuts confirm that open-weight Chinese models have reset inference economics: Ramp's vendor index shows US firms are now paying DeepSeek directly, routing data to China-hosted servers to undercut Silicon Valley pricing. Stability AI, meanwhile, released Stable Audio 3, an open-weight music generation model family trained on licensed data. The contrast is sharp. China is racing to the bottom on cost, while Europe is testing whether licensed training data can make generative media commercially viable without litigation.

Sources: SCMP, SCMP, Stability AI

Policy & Power

The US Bureau of Industry and Security issued guidance on May 31 that expands advanced AI chip export licensing requirements to Chinese-headquartered firms operating outside China, the South China Morning Post reported on June 6. The rule pushes companies toward Southeast Asian data centers and tightens the noose on offshore workarounds. On the same day, S&P Dow Jones Indices confirmed it will not waive the one-year profitability rule for index entry, blocking fast-track inclusion for SpaceX and, by extension, OpenAI and Anthropic if they list before the waiting period ends. Microsoft separately tightened human-rights controls after an inquiry found Israel's Unit 8200 violated its Terms of Service. In Europe, Mistral CEO Arthur Mensch warned that Europe has roughly two years to build sovereign AI infrastructure or risk becoming a "vassal state," and is actively exploring custom chip design.

Sources: SCMP, Ars Technica, The Guardian, Business Insider

Eastern Front

Moonshot AI is in talks to raise more than $1 billion at a $30 billion valuation, Bloomberg reported on June 8, up from $20 billion in its last round. The Beijing-based lab behind the Kimi chatbot would become China's most richly funded private AI startup after DeepSeek. The capital is not the only milestone. Cerebras revealed on May 20 that it is serving Moonshot's Kimi K2.6 model at 981 tokens per second, 6.7 times faster than the fastest GPU cloud, enabling real-time agentic coding on Cerebras wafer-scale chips. The partnership means a Chinese frontier model is running on American non-NVIDIA silicon at speeds that challenge NVIDIA's inference monopoly. Huawei's own chip milestone — completing full-parameter post-training of DeepSeek-V4-Pro on domestic Ascend 910C silicon — remains the backdrop. Chinese labs are securing both foreign inference partnerships and domestic training independence at once.

Sources: Bloomberg, Cerebras

The View

OpenAI's superapp pivot and China's price war describe opposite sides of the same transition. The chat interface is being disassembled: in the West it is being rebuilt as an agentic operating system, while in the East it is being commoditized to the point where inference costs less than a text message. Both moves are rational. OpenAI needs platform margins to justify a public valuation; Chinese labs need volume to amortize hardware built under export controls. The tension is that these strategies are incompatible inside the same market. If OpenAI charges subscription fees for agents while DeepSeek offers comparable inference at near-zero marginal cost, users will arbitrage the gap aggressively. The real test is whether OpenAI can build agentic tools sticky enough that users do not care about the price of raw tokens underneath.

The Miss

A 2026 Bain survey found that 60% of executives say their company's data and technologies are not ready for AI, even as 90% are experimenting with it. The MIT Media Lab reported in August 2025 that 95% of generative AI pilots fail because of integration gaps and learning deficits between available tools and implementation teams. The finding suggests the industry is overbuilding models for a deployment layer that does not yet exist. Coverage has focused on frontier benchmarks rather than the messy middle where most enterprise AI spending currently burns.

Source: Rest of World

Pull Quotes

"Chat is dead." — OpenAI employee, to Financial Times, June 7

"India's pragmatic approach to AI sovereignty." — Cameron Kerry, former acting US Commerce secretary, via Rest of World

"Europe has two years before becoming America's AI vassal state." — Arthur Mensch, Mistral CEO, via Business Insider

"The real question is not who builds the most capable model. It is who can make AI work inside messy, complex enterprise environments." — Ashwin Venkatesan, HFS Research, via Rest of World

Out

The real arbitrage is not between models. It is between agentic platforms that capture margins and commodity inference that erases them.

By Neo