Google TPU Veteran Joins Anthropic to Build Custom AI Chips
Anthropic is moving deeper into the semiconductor layer of the AI stack.
According to Bloomberg, the company is building an in-house chip team and has recruited Amir Salek, a former Google TPU leader, to join its compute organization and help drive its custom silicon initiative.
Salek brings unusually relevant experience to the effort. At Google, he led the TPU project from 2013 through 2022 and oversaw the research, development, and deployment of the first seven TPU generations. Earlier in his career, he spent eight years at NVIDIA, where he contributed to the development and expansion of the company’s system-on-chip business.
His appointment suggests that Anthropic is no longer viewing compute primarily as an infrastructure procurement problem. Instead, the company appears increasingly interested in controlling parts of the underlying hardware stack that powers Claude training and inference.
🧠 Anthropic Is Moving Closer to the Compute Layer #
Anthropic has historically relied on external infrastructure providers to meet the rapidly increasing compute requirements of its Claude models.
Its current strategy spans multiple hardware ecosystems, including NVIDIA GPUs, Google TPUs, and Amazon Web Services infrastructure.
This multi-vendor approach provides flexibility and reduces dependence on a single accelerator platform, but it also leaves Anthropic exposed to hardware availability, pricing, supply constraints, and the architectural limitations of third-party accelerators.
As model sizes, context windows, agent workloads, and inference volumes increase, those constraints become increasingly significant.
Compute spending is becoming a strategic variable #
Anthropic is reportedly expecting approximately $19 billion in compute expenditure during 2026.
At that scale, even relatively small improvements in utilization, memory efficiency, networking, or inference throughput can translate into substantial financial savings.
The economics also change once inference becomes a persistent production workload.
Training requires enormous bursts of compute during development cycles. Inference, by contrast, continues for as long as users interact with the model.
For a widely deployed model such as Claude, inference efficiency can therefore become a recurring operating-cost advantage.
🔧 Why Anthropic Wants Custom AI Silicon #
Building a custom accelerator gives an AI company greater control over the hardware-software interface.
Instead of adapting models to a general-purpose accelerator, the company can potentially design hardware around the computational characteristics of its own workloads.
Potential optimization targets include:
- Matrix and tensor computation
- Memory bandwidth
- SRAM capacity
- KV-cache handling
- Transformer execution
- MoE routing
- Interconnect bandwidth
- Power efficiency
- Inference scheduling
- Model-specific data movement
The objective is not necessarily to outperform NVIDIA across every workload.
A more realistic goal is to achieve better cost per token, performance per watt, and utilization for Anthropic’s own workloads.
Model-hardware co-design becomes more valuable at scale #
Modern AI models increasingly expose workload characteristics that specialized hardware can exploit.
If Anthropic knows the architecture and execution patterns of future Claude models in advance, it can theoretically build an accelerator optimized for those patterns.
This creates a tighter feedback loop:
Model architecture
↓
Workload characteristics
↓
Custom accelerator design
↓
Compiler/runtime optimization
↓
Higher utilization
↓
Lower inference cost
The more compute Anthropic consumes, the more valuable this optimization loop becomes.
👨💻 Amir Salek Brings Direct TPU Experience #
Salek’s background makes his recruitment particularly notable.
At Google, he reportedly led the TPU organization from 2013 to 2022, covering the development and deployment of the first seven generations of Google’s custom AI accelerators.
That experience spans much of the evolution of modern AI hardware, from early accelerator concepts to large-scale production systems.
Before Google, Salek spent eight years at NVIDIA and helped establish and scale its SoC business.
His combined experience covers both sides of the accelerator ecosystem: designing specialized compute hardware and building products within one of the world’s largest GPU companies.
His role at Anthropic #
After leaving Google, Salek joined Cerberus Capital Management as a Senior Managing Director.
At Anthropic, he joins the compute organization and will report to James Bradbury, Head of Compute Platform.
The appointment gives Anthropic a senior semiconductor specialist with direct experience in architecture, productization, and large-scale AI compute deployment.
That expertise is particularly relevant if Anthropic intends to move from exploratory silicon projects toward a production accelerator roadmap.
💰 Custom Chips Are a Long-Term Investment #
Custom AI silicon does not provide an immediate solution to Anthropic’s compute requirements.
Advanced accelerator development requires substantial capital and a long development cycle.
The process typically involves:
- Workload characterization
- Architecture definition
- RTL and physical design
- Verification
- Software and compiler development
- Tape-out
- Silicon validation
- Production qualification
- Packaging and system integration
- Data-center deployment
Each stage introduces technical and financial risks.
A chip can also arrive after the target model architecture has changed, potentially reducing the value of optimizations made earlier in the design cycle.
Anthropic will still need external accelerators #
For these reasons, Anthropic is unlikely to abandon NVIDIA, Google, AWS, or other infrastructure partners in the near term.
Custom silicon should instead be viewed as another layer in a broader compute strategy.
Anthropic can continue using external accelerators for flexibility and capacity while gradually introducing specialized hardware where the economics justify it.
That creates a multi-chip strategy rather than a complete replacement of third-party infrastructure.
⚔️ Anthropic Is Following a Broader AI Industry Trend #
Anthropic’s move is part of a larger shift among major AI companies.
The economics of AI increasingly favor organizations that can control more of their compute stack.
Google developed TPUs specifically to support its large-scale AI workloads.
Apple has spent years building tight hardware-software integration around its custom silicon for client devices.
Microsoft has also developed its own AI accelerator technology.
OpenAI has reportedly pursued a custom chip initiative known as Jalapeño, in partnership with Broadcom and Celestica, with a focus on AI inference workloads.
Although the technical approaches differ, the strategic objective is similar: reduce the cost and dependency associated with running increasingly expensive AI workloads at scale.
⚡ Inference Is Becoming the New Compute Battleground #
Training has traditionally received most of the attention in AI infrastructure discussions.
However, inference economics are becoming increasingly important.
Training a new model may require enormous amounts of compute over a limited development period. Once deployed, the model can generate inference workloads continuously across millions of user interactions.
This creates a fundamentally different cost profile.
A specialized inference accelerator could potentially improve:
- Tokens per watt
- Tokens per dollar
- Memory utilization
- Batch efficiency
- Latency
- Power consumption
- Data-center density
For an AI company operating at massive inference volume, these improvements can compound rapidly.
Agentic AI increases the pressure #
The emergence of AI agents could make inference efficiency even more important.
An agent may invoke a model repeatedly while performing a task rather than generating a single response.
A workflow could involve:
User request
↓
Model inference
↓
Tool call
↓
New context
↓
Model inference
↓
Another tool call
↓
Additional inference
As the number of inference cycles per task increases, hardware efficiency becomes increasingly important to the economics of the overall service.
Custom accelerators optimized for these recurring workloads could therefore provide strategic value beyond conventional chatbot serving.
🏭 The AI Compute Supply Chain Is Being Rebuilt #
The move toward custom silicon is also changing the competitive structure of the semiconductor industry.
NVIDIA remains the dominant provider of AI accelerators, but major AI companies increasingly have incentives to diversify their hardware sources.
That does not necessarily mean replacing NVIDIA.
Instead, the industry is moving toward a more heterogeneous model involving:
- General-purpose AI GPUs
- Cloud-provider accelerators
- Custom inference ASICs
- Specialized networking hardware
- Advanced packaging
- High-bandwidth memory
- Custom interconnects
- Software-defined compute platforms
This gives large AI companies more control over their long-term capacity planning.
Hardware supply is becoming a strategic asset #
The AI industry has already demonstrated that accelerator availability can constrain model development and deployment.
If an AI company controls more of its hardware roadmap, it can potentially align accelerator production with its own expected compute demand.
That could become especially important as model providers compete for increasingly scarce advanced packaging capacity, HBM, networking components, and leading-edge manufacturing capacity.
🧩 The Real Advantage Is Hardware-Software Co-Design #
A custom chip is valuable only when the surrounding software stack can exploit it.
Anthropic would therefore need to develop more than an accelerator.
A successful platform would likely require:
- Compiler support
- Kernel libraries
- Runtime systems
- Distributed inference software
- Model graph optimization
- Memory management
- Scheduling
- Profiling tools
- Developer tooling
This is one reason Salek’s TPU experience is potentially significant.
Google’s TPU strategy succeeded not simply because it built custom silicon, but because the hardware was integrated into a broader software and infrastructure stack.
Anthropic faces the same fundamental challenge.
🌐 Anthropic Could Be Becoming an AI Infrastructure Company #
The strategic implications extend beyond chip design.
Anthropic began primarily as an AI model company, with Claude at the center of its product strategy.
But as compute becomes a larger portion of the cost structure, controlling infrastructure becomes increasingly valuable.
The progression can be viewed as:
AI models
↓
Model serving
↓
Compute infrastructure
↓
Custom accelerators
↓
Hardware-software co-design
Each layer gives the company greater control over performance, cost, and supply.
Anthropic’s recruitment of a former TPU leader suggests that the company is exploring how far down this stack it should go.
🔮 AI Competition Is Moving Beyond Model Benchmarks #
The next stage of AI competition may not be determined exclusively by which company produces the strongest model.
The more difficult question is who can operate that model most efficiently at massive scale.
That means competition increasingly spans:
- Model architecture
- Training efficiency
- Inference cost
- Accelerator design
- Memory systems
- Networking
- Data centers
- Power efficiency
- Compiler technology
- Supply-chain control
Anthropic’s decision to recruit a veteran of Google’s TPU program is therefore more than a senior engineering hire.
It signals that the company is taking the compute layer itself seriously as a competitive advantage.
NVIDIA, Google, AWS, Broadcom, and emerging custom-silicon providers will continue to play important roles, but major AI labs increasingly have incentives to build differentiated hardware capabilities of their own.
The long-term AI race may consequently become a race not only to build smarter models, but to build the most efficient machine for running them.