Scale & Strategy
together with
IBM
This is Scale & Strategy, the lighthouse guiding you home through choppy AI seas.
Here’s what we’ve got for you today:
Chinese startup unveils homegrown AI chip to challenge Nvidia
Meta backs down after AI photo feature sparks user revolt
Chinese startup unveils homegrown AI chip to challenge Nvidia
China just revealed another contender in the AI chip race, and this one is taking a very different approach.
Startup Dongfang Suanxin (DFSX) emerged from stealth this week with the launch of the DF1000, an AI chip built entirely on a domestic supply chain. The company also outlined an aggressive roadmap, with the DF2000 expected later this year and a third-generation chip planned for 2027.
DFSX was founded in 2024 but intentionally stayed quiet while developing its technology. Following its latest funding round, the company is reportedly valued at roughly $1.8B, with backing from state-linked investors and industrial funds, including a venture firm co-founded by Alibaba founder Jack Ma.
Instead of trying to match Nvidia process node for process node, DFSX is attacking a different bottleneck.
The company uses mature 14nm manufacturing rather than cutting-edge fabrication. To make up the difference, it stacks custom memory directly on top of the compute layer, dramatically increasing memory bandwidth without relying on advanced High Bandwidth Memory (HBM), one of the biggest choke points created by U.S. export restrictions.
DFSX claims the DF1000 can match some mainstream Western chips on certain AI inference workloads despite using far older manufacturing. Training performance still trails today’s frontier accelerators, though the company says the upcoming DF2000 will narrow that gap.
The architecture also differs from China’s current AI strategy.
While companies like Huawei compensate for manufacturing limitations by connecting massive clusters of chips together, DFSX focuses on reducing latency inside a single chip by placing memory closer to compute and dynamically reconfiguring hardware around the software it’s running.
The technology builds on roughly two decades of research led by Tsinghua University professor and government advisor Wei Shaojun.
The bigger story isn’t whether the DF1000 beats Nvidia today. It almost certainly doesn’t.
The more important question is whether China can build a competitive AI hardware ecosystem without access to Western manufacturing. If companies like DFSX can deliver “good enough” performance using domestic fabs and novel architectures, U.S. export controls become less about stopping China’s AI progress and more about changing the path it takes to get there. That’s a much different outcome than simply freezing the country’s semiconductor ambitions.
Guiding AI with project-level rules
AI agents are reshaping development, reportedly generating an average of 48% of code for surveyed organizations. But with that shift comes growing concern: 55% of engineering leaders surveyed are concerned about losing shared understanding of how their codebase evolves, and 39% are worried about shipping with confidence.
The issue is not that AI produces bad code. It is that each agent makes different decisions on frameworks, testing and patterns. Over time, this can create inconsistency that’s harder to review and maintain. Project-level rules provide a structured way to address this, encoding conventions and standards directly into workflows so AI-generated code remains aligned with how teams build and maintain software.
See how to help keep AI code consistent
Meta backs down after AI photo feature sparks user revolt
Meta’s new AI image tools barely made it out the door before users forced the company into a retreat.
Last week, Meta Superintelligence Labs launched Muse Image, its new image generation model designed to compete with tools like Google’s Nana Banana and ChatGPT Images 2.0. The pitch was simple: generate high-quality images directly inside Instagram and WhatsApp, with Facebook and Messenger coming next.
Buried in the announcement, however, was a feature that quickly stole the spotlight.
Meta allowed users to generate AI images based on content from public Instagram accounts by simply mentioning them with an @. The goal was to create images “rooted in your world.” The problem was that the feature was enabled by default, many users didn’t realize they had been opted in, there was no notification when someone generated an image of them, and there wasn’t a clear way to remove existing images.
The backlash was immediate.
Within days, Meta pulled the feature.
“Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way,” the company said in a statement.
The reversal is one of the clearest examples yet of public pressure changing an AI product after launch. It also highlights how much more sensitive AI becomes when it moves beyond generating landscapes and logos into using people’s identities.
“You can reduce the risk. Avoid posting high-resolution close-ups of faces or hands. Strengthen privacy settings so only trusted people can view your content,” said Trua CEO Raj Ananthanpillai, author of The Trust Crisis.
Meta won’t be the last company to run into this problem.
Every major AI lab is building products that blur the line between public content and training data, or between public content and AI-generated derivatives. As AI becomes embedded inside cameras, glasses, messaging apps, and social networks, the debate shifts from what models are trained on to what they’re allowed to generate in real time.
The technical challenge isn’t making these systems more capable anymore. It’s figuring out where users draw the line before regulators do it for them.
Was this email forwarded to you?
That’s it for today and as always It would mean the world to us if you help us grow and share this newsletter with other operators.
Our mission is to help as many business operators as possible, and we would love for you to help us with that mission!