Pang joined Meta in mid-2025 after leaving Apple, where he focused on AI model infrastructure.

At Meta, Pang oversaw key systems supporting next-generation AI models. His exit comes after months of recruitment by OpenAI, and sources said he left Meta last week.

Meta and OpenAI did not immediately respond to requests for comment.

The development reveals the scale of investment companies are making in AI talent. Reports indicate his compensation at Meta was more than $200 million over multiple years.

His work at Meta was a major one in developing the company’s next-generation AI systems, part of its long-term strategy to compete with firms including OpenAI, Anthropic, Apple, and Google.

For OpenAI, Ruoming Pang strengthens its infrastructure and ability to scale. His expertise will support the training of larger multimodal models and more advanced AI workflows.

The hire occurs as governments strengthen review of AI development. Canada, for example, has recently urged OpenAI to increase safety measures or face regulatory consequences.

The arrangement, finalised in May 2025 after months of negotiations, went live this quarter, making OpenAI one of Google Cloud’s biggest customers to date.

Sundar Pichai, CEO of Google, confirmed the deal during Alphabet’s second-quarter earnings call, stating:

“We are very excited to be partnering with them on Google Cloud. Google Cloud is an open platform, and we have a strong history of supporting great companies, startups, AI labs, etc. So super excited about our partnership there on the cloud side, and we look forward to investing more in that relationship and growing that.”

OpenAI’s demand for compute has surged to historic levels, with projections indicating over one million GPUs will be online by the end of 2025. Its longer-term goal involves scaling to 100 million GPUs over the next decade. 

This explosive growth has outpaced Microsoft Azure’s capacity, the company’s primary cloud partner, forcing OpenAI to look elsewhere.

Google Cloud, once kept at arm’s length by OpenAI due to its rivalry with Microsoft and shared ambitions in AI, now enters the scene with a compelling offer: access to Nvidia’s top-tier H100 and GB200 GPUs, custom-built TPUs, and a mature infrastructure trusted by leading AI research outfits including Anthropic, Fei-Fei Li’s World Labs, and Ilya Sutskever’s Safe Superintelligence.

The exclusivity OpenAI once held with Microsoft ended quietly in January 2025. Under the new “right of first refusal” framework, OpenAI gained the freedom to diversify its backend. 

The company’s operations now span data centres in Oregon, Iowa, Frankfurt, Milan, Singapore, and Tokyo, optimised for low-latency inference and strict sovereign compliance. Workloads are split across Azure and Google Cloud using orchestration tools like Kubernetes, Anthos, and Istio.

This development reveals the pressure OpenAI faces in scaling its services while maintaining uptime, responsiveness, and sustainability commitments. Both firms have pledged to power AI workloads with 100% carbon-free energy, and will publish energy transparency metrics in line with grid-optimised strategies.

While OpenAI remains the most direct threat to Google’s crown jewel, Search, the economic logic behind this partnership is clear. Google Cloud posted a $13.62 billion revenue haul in Q2 2025, up 32% year-on-year. 

Much of that growth comes from AI workloads, prompting Alphabet to raise its capital expenditure target for the year to $85 billion, noting “strong and growing demand” across both AI and cloud services.

Still, the optics are awkward. OpenAI’s ChatGPT product has chipped away at Google’s dominance in Search, forcing the tech giant to hasten development on its own generative AI offerings like Gemini. 

That chatbot now reaches 450 million monthly users, while AI Overviews on Search reportedly sees 2 billion monthly users. But the monetisation strategy behind those products is not well explained.

Some analysts are comparing this move to Google’s early relationship with Yahoo, when it quietly powered Yahoo Search in its early days, only to displace it as the internet’s front page. 

Hopefully, history will not repeat itself.

Initially aiming for $2 billion, the company has now secured greater backing from investors, including Lightspeed Venture Partners, General Catalyst, Bessemer Venture Partners, and Abu Dhabi-based MGX. 

This latest investment will bring Anthropic’s total funding to approximately $18 billion.

Even with its growing revenue, which now stands at around $1.2 billion annually, Anthropic is not yet profitable. The company plans to use the funds to further its AI research and development efforts, particularly in creating more advanced models.

Anthropic has also launched Claude 3.7 Sonnet, a new AI model that allows users to choose between instant responses and deeper, more complex reasoning. Unlike traditional AI chatbots that force users to pick between models with varying capabilities, Claude 3.7 Sonnet integrates both functionalities in a single system.

According to Anthropic, the model’s unique feature is its ability to “think” for as long as needed. This approach is just like human reasoning, where some questions require quick answers while others demand more thoughtful consideration. 

“Similar to how humans don’t have two separate brains for questions that can be answered immediately versus those that require thought,” the company stated in a blog post, “we regard reasoning as simply one of the capabilities a frontier model should have, to be smoothly integrated with other capabilities, rather than something to be provided in a separate model.”

Pricing and Performance

The Claude 3.7 Sonnet model is priced at $3 per million input tokens and $15 per million output tokens, making it more expensive than some competitors, such as OpenAI’s o3-mini and DeepSeek’s R1. However, Anthropic says that its hybrid nature—combining standard chatbot functions with deep reasoning—justifies the cost.

Performance tests show Claude 3.7 Sonnet surpassing previous versions. On SWE-Bench, a benchmark for real-world coding tasks, it achieved an accuracy of 62.3%, outperforming OpenAI’s o3-mini model, which scored 49.3%. 

Another test, TAU-Bench, measured AI interaction with users in a simulated retail environment, where Claude 3.7 Sonnet scored 81.2%, ahead of OpenAI’s o1 model at 73.5%.

Again, Anthropic claims to have greatly improved the model’s ability to distinguish between harmful and harmless prompts, reducing unnecessary refusals by 45% compared to its predecessor, Claude 3.5 Sonnet.

Alongside Claude 3.7 Sonnet, Anthropic is launching a research preview of Claude Code, an AI-powered coding assistant. This tool allows developers to execute tasks directly from their terminal, such as modifying codebases, debugging errors, and even pushing updates to GitHub.

Illustrating its abilities, an Anthropic representative showed how Claude Code could break down a project’s structure and suggest improvements using plain English commands. Initially, access to Claude Code will be limited and granted on a first-come, first-served basis.

While Anthropic has taken a measured, safety-focused approach in the past, this launch seems to be a more aggressive innovation. However, OpenAI is reportedly preparing to release its own hybrid AI model in the coming months, bringing even fiercer competition in the space.