Steven Sinofsky on Apple at 50, Microsoft, and the Future of Computing
Steven Sinofsky, former Microsoft Windows division president, discusses NVIDIA's RTX Spark Super Chip announcement at Computex, the evolving AI-native computing landscape, and the ongoing tension between Apple and Microsoft's platform strategies. He argues that local AI compute will inevitably displace cloud-based token costs, and critiques Microsoft's approach of maintaining backward compatibility rather than embracing a clean break with legacy Windows architecture.
Summary
The conversation opens with discussion of NVIDIA's major Computex announcement: the RTX Spark Super Chip, an ARM CPU mated with NVIDIA parallel processing graphics into a unified system-on-chip with a new memory architecture. Sinofsky contextualizes this as NVIDIA entering the mainstream PC chip business, drawing parallels to a similar—though far smaller-scale—moment in 2011 when he was at Microsoft and NVIDIA was listed as a Surface partner. He notes Jensen Huang has achieved a Taylor Swift-like cultural prominence in tech, with this Computex keynote reaching mainstream visibility unprecedented for that traditionally inside-baseball trade show.
Sinofsky's central thesis on AI computing is that the current model of paying per token for cloud-based AI inference is a temporary constraint that will inevitably migrate to local devices, following the historical pattern of every scarce computing resource eventually becoming free and local. He uses the example of people running stacks of Mac minis to run local agents as evidence of early adoption of this trend, driven by both cost and privacy concerns. He predicts that within six to nine months, AI-native local compute will become the dominant paradigm.
On the question of Apple vs. Windows for AI-native computing, Sinofsky highlights a critical near-term unknown: what Apple will announce at WWDC regarding CUDA API support. He explains the long history of NVIDIA being an 'outsider' to both Windows and Mac platforms—always an add-on requiring separate drivers—and notes that Microsoft has now signaled CUDA will be supported on Spark devices, though the implementation details remain unclear. He sees Apple's decision on CUDA as equally pivotal and unresolved.
Sinofsky is openly critical of Microsoft's strategic direction with the Spark announcement. He argues that by emphasizing backward compatibility—running all legacy Win32 apps on ARM—Microsoft is repeating a mistake he tried to avoid when designing the original Surface. His original vision for Surface was to introduce a platform discontinuity, breaking from legacy Windows to create a sealed, mobile-first, malware-resistant experience. That strategy was abandoned after his departure, and he sees Microsoft once again choosing the backward-looking path by promising that NVIDIA Spark laptops will run every existing Windows application. He argues consumers actually want a Mac/phone-like sealed experience even if they don't articulate it, and that legacy compatibility is a feature that tests well but ultimately degrades the user experience over time.
On the Dell XPS 13 vs. MacBook Neo comparison, Sinofsky praises both machines as quality products for general consumers but dismisses the framing of one as a 'killer' of the other. He expresses concern about the XPS 13 shipping with 8GB RAM as a base configuration, calling it insufficient for a reasonable Windows experience without significant technical intervention. He personally recommends the Dell XPS 13 at 16GB as his go-to PC recommendation for friends and family, and notes the Surface lineup—while something he can't be objective about—has drifted from his original vision into a niche Intel-based product line.
Key Insights
- Sinofsky argues that paying per token for cloud AI inference is a temporary constraint that will inevitably resolve itself through local on-device compute, following the historical pattern of every scarce paid computing resource eventually migrating to the device and becoming free.
- Sinofsky claims Microsoft's decision to support all legacy Win32 apps on NVIDIA Spark ARM devices repeats the same backward-looking mistake that undermined the original Surface vision, arguing that consumers would actually prefer a sealed, registry-free, malware-resistant experience—they just don't know how to ask for it.
- Sinofsky identifies Apple's forthcoming WWDC announcements on CUDA API support as a critical and currently unknown variable that will significantly shape which platform becomes dominant for AI-native computing, noting the decision could range from native support to a thunking layer to App Store distribution.
- Sinofsky contends that NVIDIA has historically been an outsider add-on to both Windows and Mac—never a first-class OS citizen—and that the current Spark announcement represents a genuine architectural shift, though its significance depends entirely on how deeply Microsoft integrates CUDA into the OS versus treating it as a downloadable driver.
- Sinofsky reveals that the original Surface program was deliberately designed as a platform discontinuity—intentionally breaking Win32 compatibility on ARM to force an ecosystem shift to new APIs—and that the x86 Surface was only created as an 'objection handler,' a strategy he views as having been abandoned prematurely after his departure from Microsoft.
Topics
Full transcript available for MurmurCast members
Sign Up to Access