Listen to this Post
A New Chapter for Anthropic
The artificial intelligence race is no longer being fought only inside research labs. Increasingly, it is being fought inside data centers, semiconductor factories, power stations, networking facilities, and the enormous supply chains that keep modern AI systems alive.
Anthropic, the company behind Claude, appears to be taking another major step in that direction.
The company has hired Amir Salek, a semiconductor executive who previously helped build Alphabet’s Google Tensor Processing Unit (TPU) program, as part of its compute organization. Salek’s arrival is significant because Anthropic is already one of the largest buyers of AI computing capacity, relying on a combination of NVIDIA GPUs, Google TPUs, and Amazon’s Trainium processors.
The message behind the appointment is difficult to miss: Anthropic increasingly wants to have a say in the silicon powering its future.
That does not necessarily mean Anthropic is about to become a conventional chip manufacturer. Instead, the company could be moving toward a more vertically integrated AI infrastructure strategy in which its engineers have greater control over processor architecture, memory systems, networking, inference performance, and the economics of running Claude at enormous scale.
The timing could hardly be more important.
AI models are becoming larger, inference workloads are exploding, and the cost of operating frontier systems is becoming one of the biggest strategic constraints facing the industry. Anthropic itself has dramatically expanded its compute commitments. The company has announced major expansions involving Google TPUs and Broadcom, while continuing to work with AWS and NVIDIA.
Now, with a veteran of
Who Is Amir Salek?
Salek is not simply another technology executive joining Anthropic.
He is closely associated with the development of Google’s TPU strategy, the custom AI accelerator program that became one of the most important alternatives to NVIDIA’s dominance in machine-learning hardware.
According to the original report, Salek helped establish Google’s custom-chip program and led the TPU business through 2022, overseeing the delivery of the first seven generations of Google’s specialized processors.
That experience is particularly relevant to Anthropic because Google’s TPU story demonstrates why custom silicon can matter so much to an AI company.
Rather than relying exclusively on general-purpose accelerators, Google designed processors around the computational characteristics of machine-learning workloads. That allowed the company to optimize hardware, software, networking, memory, and data-center architecture as a unified system.
Anthropic could now be attempting to apply some of those lessons to Claude.
Why Anthropic Needs More Control
The biggest problem facing frontier AI companies is not simply model intelligence.
It is compute.
A company can design an extraordinary model, but if it cannot obtain enough processors, electricity, networking equipment, memory, cooling capacity, and data-center space, that model cannot serve billions of requests.
Anthropic already understands this reality.
The company has been expanding its hardware footprint through multiple suppliers rather than placing all of its infrastructure bets on one platform. Anthropic says it runs Claude across AWS Trainium, Google TPUs, and NVIDIA GPUs, giving it greater flexibility and resilience.
That diversification is strategically valuable.
But custom silicon could take the strategy one step further.
Instead of merely asking hardware companies to provide processors, Anthropic could increasingly influence how those processors are designed for Claude’s particular workloads.
The Economics Behind the Silicon Race
AI inference is extraordinarily expensive.
Every time a user asks Claude a question, generates code, analyzes a document, summarizes a meeting, or performs a complex agentic task, computing resources are consumed.
At small scale, the cost may look insignificant.
At global scale, it becomes enormous.
This is why inference efficiency has become one of the central battles in AI infrastructure.
A processor that delivers the same useful output with less power, lower memory traffic, better utilization, or lower latency can potentially save an AI company enormous amounts of money.
The advantage does not need to come from a dramatic improvement in raw performance.
Even a modest efficiency gain multiplied across millions or billions of requests can become financially transformative.
Anthropic Is Already Diversifying Its Hardware
The Salek appointment should therefore be viewed as part of a much larger infrastructure strategy rather than an isolated hiring decision.
Anthropic has expanded its relationship with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to begin coming online in 2027. The company has also emphasized that its hardware strategy includes AWS Trainium, Google TPUs, and NVIDIA GPUs.
That is a remarkable position.
Anthropic is effectively building an AI platform that does not depend on one processor architecture.
The company can potentially match different workloads with different hardware.
Training may favor one architecture.
Inference may favor another.
Large-scale agentic workloads could eventually require yet another configuration.
Custom silicon gives Anthropic the possibility of optimizing around these differences.
The Fractile Connection
Another important piece of the story is Fractile, the British AI-chip startup developing specialized inference hardware.
Reports indicate that Anthropic has reached an initial agreement involving roughly $250 million worth of Fractile chips, with the possibility of expanding the relationship. The chips are not expected to be ready for deployment until 2027.
Fractile itself has positioned its technology around the difficult economics of AI inference, arguing that latency and cost are increasingly limiting the usefulness of advanced AI systems. The company raised $220 million in May 2026 to accelerate development of its hardware and systems.
This is important because it shows that
Instead, the company could build an ecosystem.
It could use NVIDIA for certain workloads.
Google TPUs for others.
AWS Trainium for additional capacity.
Fractile for specialized inference.
And eventually, potentially, Anthropic-designed silicon for workloads where it believes deeper customization offers a competitive advantage.
The OpenAI Comparison
Anthropic is not alone.
OpenAI has already demonstrated how seriously leading AI companies are taking custom hardware.
In June 2026, OpenAI and Broadcom unveiled Jalapeño, an inference processor designed specifically around large language model workloads. OpenAI said the chip was developed from scratch with Broadcom and is intended to form part of a multigenerational compute platform.
The significance is enormous.
OpenAI is no longer positioning itself purely as an AI model company.
It is moving deeper into the infrastructure stack.
Anthropic appears to be moving in a similar direction.
The competitive question is therefore changing from:
Who has the best AI model?
to:
Who can build, power, and operate the most efficient AI machine?
Why Google TPU Experience Matters
Salek’s background could be especially useful because Google spent years solving exactly this problem.
The TPU program required coordination between chip designers, compiler engineers, machine-learning researchers, data-center architects, networking specialists, and software teams.
Custom AI silicon is not simply about designing a faster processor.
The entire software stack must understand the processor.
Models must be optimized for it.
Kernels must be written for it.
Memory must be carefully managed.
Networking must be designed around it.
Schedulers must understand its characteristics.
And data centers must be capable of deploying it efficiently.
Someone who has already experienced this process at Google’s scale brings a rare combination of technical and organizational knowledge.
The Real Goal May Be Inference
One of the most interesting possibilities is that Anthropic’s hardware ambitions could initially focus heavily on inference rather than training.
Training frontier models requires immense computational resources, and the workloads are extraordinarily demanding.
But inference happens continuously after a model has been trained.
Every customer interaction becomes an inference workload.
That makes inference efficiency particularly attractive.
If Anthropic can reduce the cost of serving Claude while maintaining quality and speed, it could potentially improve margins without requiring a corresponding increase in customer prices.
This becomes even more important as AI agents become capable of taking dozens, hundreds, or potentially thousands of computational steps to complete complex tasks.
The Agentic AI Factor
The rise of agentic AI could make custom inference hardware even more valuable.
A traditional chatbot may generate a relatively short response.
An AI agent can reason, call tools, inspect files, write code, execute tasks, evaluate results, and repeat the process.
That means one user request can generate a large number of model calls.
The computational burden can therefore multiply rapidly.
Hardware designed around these repeated inference patterns could eventually provide a major advantage.
This is where
Why NVIDIA Still Matters
None of this means NVIDIA is suddenly irrelevant.
Quite the opposite.
NVIDIA remains one of the most important suppliers in the AI ecosystem, and custom accelerators are unlikely to replace GPUs across every workload overnight.
General-purpose AI accelerators provide enormous flexibility.
They benefit from mature software ecosystems.
They support a huge variety of workloads.
And
The likely future is therefore not “custom chips replace NVIDIA.”
It is custom chips plus GPUs plus specialized accelerators.
The AI companies that manage that mixture most effectively could gain a major advantage.
Anthropic’s Full-Stack Ambition
The deeper story is that Anthropic increasingly appears to be moving toward full-stack AI infrastructure.
At the top is Claude.
Below Claude are model-serving systems, inference software, compilers, kernels, networking, memory, processors, and data centers.
At the bottom are electricity, cooling, construction, semiconductor supply chains, and physical infrastructure.
Every layer affects the others.
A more efficient chip can reduce power consumption.
Lower power consumption can reduce operating costs.
Lower costs can allow more inference.
More inference can support more capable products.
More capable products can attract more customers.
More customers create greater demand for compute.
That creates a feedback loop.
The Data-Center Arms Race
The semiconductor story also cannot be separated from the data-center race.
Anthropic has been signing large infrastructure agreements at a rapid pace because frontier AI models require enormous amounts of computing capacity.
Its expansion with Google and Broadcom is one example. The company has also been building its relationship with AWS, its primary cloud provider and training partner.
Meanwhile, the wider industry is committing extraordinary sums to AI infrastructure.
Recent financing activity around Broadcom and AI data-center projects illustrates how capital-intensive this market has become.
The result is a new type of technology arms race.
Companies are no longer simply racing to hire researchers.
They are racing to secure electricity, chips, networking equipment, cooling systems, construction capacity, and semiconductor expertise.
Why Custom Silicon Could Reduce Risk
There is another reason Anthropic may be interested in hardware.
Supply-chain resilience.
If one supplier becomes constrained, Anthropic needs alternatives.
If NVIDIA capacity becomes difficult to obtain, Google TPUs and AWS Trainium provide options.
If specialized inference chips become commercially viable, companies such as Fractile could provide another source.
And if Anthropic eventually controls part of its own silicon roadmap, it gains even more strategic independence.
That does not eliminate supply-chain risk.
But it can reduce dependence on any single supplier.
The Hidden Advantage: Better Optimization
Custom silicon also allows an AI company to design hardware around its own software.
That is potentially more important than simply having a proprietary processor.
Imagine Anthropic knows exactly how future Claude models will generate tokens, access memory, call tools, use long contexts, and interact with external systems.
It could design hardware specifically around those behaviors.
That is the ultimate promise of vertical integration.
The model informs the software.
The software informs the hardware.
The hardware improves the software.
And the cycle repeats.
Why This Could Change AI Competition
For years, the dominant AI narrative focused on model benchmarks.
Then attention shifted toward data.
Then compute.
Now the industry is moving toward infrastructure efficiency.
The next generation of competition may therefore be determined by something much less visible to ordinary users: how many useful AI operations a company can produce per dollar and per watt.
That is where custom silicon becomes strategically powerful.
What Undercode Say:
1. The Hardware War Has Arrived
Anthropic hiring a Google TPU veteran is more than a personnel announcement. It reflects the broader transformation of AI companies into infrastructure companies.
2. Claude Needs Physical Infrastructure
Claude may look like software to users, but behind every response sits an enormous physical machine consuming electricity and compute capacity.
3. Silicon Is Becoming Strategic
The processor is no longer just a component purchased from a supplier. For frontier AI companies, it can become a strategic layer of the business.
4. Google Already Proved the Model
Google demonstrated that custom AI accelerators can become a fundamental part of a company’s AI strategy rather than an experimental side project.
5. Salek Brings Valuable Experience
Someone involved in multiple generations of TPU development understands the difficulties of turning AI research requirements into practical silicon.
6. Anthropic Is Taking a Diversification Approach
The company does not appear to be betting everything on one processor.
Instead, it is building relationships across NVIDIA, Google, AWS, Broadcom, and emerging chip companies.
7. That Strategy Makes Sense
The AI hardware market changes too quickly for a single architecture to remain ideal for every workload.
8. Different Workloads Need Different Chips
Training, inference, reasoning, retrieval, agentic execution, and specialized AI workloads can have very different hardware requirements.
9. Inference May Become the Biggest Battlefield
As AI becomes a daily service, inference could consume an enormous portion of total computing resources.
10. Agents Increase the Pressure
An AI agent may need to invoke models repeatedly during one task.
That means the computational cost of a single customer interaction can grow dramatically.
11. Efficiency Will Matter More
If two companies provide similar AI quality but one can serve the same workload for substantially less money, the efficiency advantage could become decisive.
12. Power Is Part of the Equation
AI chips are ultimately limited by electricity, cooling, and data-center capacity.
A more efficient accelerator can therefore provide advantages far beyond the processor itself.
13. Memory Is Critical
Modern AI workloads frequently move enormous quantities of data.
Reducing unnecessary memory movement can improve both performance and energy efficiency.
14. Networking Matters Too
Large AI systems are not isolated processors.
Thousands of accelerators may need to communicate rapidly.
That makes networking architecture a crucial component of overall performance.
15. Software Cannot Be Ignored
A brilliant chip without a strong software ecosystem can become a very expensive experiment.
Anthropic would need compilers, kernels, runtime systems, debugging tools, schedulers, and developer infrastructure around any proprietary silicon.
16. That Is Why
The challenge is not just designing silicon.
It is building an entire ecosystem around that silicon.
17. OpenAI Is Following the Same Direction
OpenAI’s Jalapeño processor demonstrates that the shift toward custom AI hardware is becoming an industry-wide phenomenon.
18. Google Is Still a Major Force
Ironically,
Anthropic has announced major expansions involving
19. AWS Remains Important
Anthropic also continues to rely heavily on
20. NVIDIA Is Not Going Away
The custom-chip movement should not be mistaken for the collapse of GPU demand.
General-purpose accelerators remain extraordinarily valuable.
21. The Future Is Hybrid
The likely winning architecture is not one processor.
It is a portfolio of processors optimized for different tasks.
22. Fractile Is Particularly Interesting
The reported Anthropic-Fractile relationship suggests that specialized inference hardware could become an important part of this portfolio.
23. Timing Is Everything
Fractile’s hardware is reportedly aimed at future deployment, meaning Anthropic is making infrastructure decisions years ahead of actual production requirements.
24. That Requires Long-Term Thinking
Frontier AI infrastructure cannot be built in a few weeks.
Chip development, manufacturing, validation, deployment, and data-center integration all require significant lead time.
25. Anthropic Is Planning Ahead
The Salek hire suggests that Anthropic is thinking about hardware requirements beyond the current generation of Claude.
26. Future Models May Be Hardware-Aware
Eventually, model architectures may be designed with particular accelerator characteristics in mind.
27. That Could Create a Competitive Moat
If
28. It Could Also Lower
Better hardware utilization could eventually reduce the amount of computing required for each useful response.
29. Lower Costs Could Accelerate Adoption
If inference becomes cheaper, Anthropic could make more powerful Claude capabilities available to more customers.
30. Enterprise AI Makes This More Important
Businesses often care as much about reliability, latency, security, and operating cost as they do about benchmark scores.
31. Reliability Requires Infrastructure
A brilliant model that becomes unavailable during peak demand is not useful to an enterprise.
32. Capacity Is Competitive Advantage
The ability to guarantee compute capacity can become part of the product itself.
33. Silicon Could Become a Strategic Asset
In the same way that cloud infrastructure became strategically important, AI accelerators may become core corporate assets.
34. Capital Requirements Will Rise
Custom hardware is expensive.
Chip design teams, verification, manufacturing, packaging, networking, and deployment all require significant investment.
- Not Every AI Company Can Do This
Only companies operating at enormous scale are likely to justify the expense of deep custom-silicon development.
36. Anthropic Is Reaching That Scale
The
37. The Chip Could Become the
Users may never know which processor answered their question.
But that processor could determine how fast, affordable, and reliable the experience becomes.
- The Real Competition Is Cost Per Intelligence
The industry may eventually stop obsessing over raw FLOPS and focus more heavily on how much useful reasoning can be delivered per dollar and per watt.
39. Anthropic Is Preparing for That Future
The Salek appointment fits neatly into this transition from model-centric competition to full-stack AI competition.
40. The Bigger Story Is Just Beginning
Anthropic may not immediately unveil a chip carrying its own logo.
But the company is clearly moving closer to the machinery that makes Claude possible.
And in the next phase of AI, controlling that machinery could matter almost as much as controlling the model itself.
Deep Analysis: What Anthropic Could Gain From Custom AI Silicon
Hardware-Software Co-Design
The biggest technical opportunity is hardware-software co-design.
A conventional accelerator is built to support a broad range of workloads.
A custom accelerator can instead be designed around the specific operations that Anthropic expects Claude to perform most frequently.
That could include matrix operations, attention mechanisms, memory access patterns, token generation, routing, batching, and increasingly complex agentic workflows.
Inference Optimization
Inference is particularly attractive because it occurs continuously.
A custom inference processor could prioritize low latency, high utilization, predictable memory access, and efficient token generation rather than attempting to optimize for every possible machine-learning workload.
OpenAI’s Jalapeño strategy provides a useful industry comparison: its processor is explicitly designed around LLM inference and is intended for deployment at gigawatt scale over multiple generations.
Memory Efficiency
Memory movement can become one of the most expensive parts of AI computing.
A processor that reduces unnecessary movement between compute units and memory can potentially reduce both latency and electricity consumption.
For long-context models, this becomes even more important.
Networking Efficiency
Large AI deployments depend on high-speed communication between processors.
If thousands of accelerators are working together, network bottlenecks can waste expensive compute resources.
A future Anthropic architecture could therefore treat networking as an integral part of the accelerator rather than an external consideration.
Compiler Optimization
Custom silicon would also require an aggressive compiler strategy.
The
The basic concept can be visualized with commands such as:
Inspect CPU architecture and available instruction sets
lscpu
Inspect PCI devices, including accelerators and high-speed hardware
lspci | grep -Ei nvidia|amd|accelerator|network
Inspect NVIDIA GPU utilization on systems using NVIDIA hardware
nvidia-smi
Monitor GPU activity continuously
watch -n 1 nvidia-smi
These commands do not represent
Measuring the Real Bottleneck
A serious hardware strategy also requires profiling.
An accelerator that looks extremely fast on paper can underperform if the surrounding software cannot keep it busy.
Engineers therefore need to measure utilization, memory bandwidth, latency, communication overhead, and power consumption.
A simplified Linux workflow might begin with:
uname -a
lscpu
free -h lsblk lspci
The objective is not simply to ask, “How powerful is this processor?”
The better question is:
How much useful AI work does the entire system produce per dollar, per second, and per watt?
The Importance of Power
Power may ultimately become one of the strongest arguments for custom silicon.
Anthropic can purchase more processors, but those processors still need electricity.
As AI data centers grow, power availability becomes a physical constraint.
A processor that performs the same workload using significantly less electricity can therefore provide an infrastructure advantage even if its raw computational performance is not dramatically higher.
Custom Chips Are Not a Magic Solution
There are serious risks.
Chip development is expensive.
Manufacturing delays can destroy schedules.
Software ecosystems take years to mature.
A custom accelerator can become obsolete if model architectures change.
And an architecture optimized for today’s Claude workloads might be poorly suited to tomorrow’s models.
This is why
The Most Likely Strategy
The most realistic outcome is probably a hybrid architecture.
Anthropic will continue using NVIDIA GPUs.
It will continue expanding Google TPU capacity.
It will continue working with AWS and Trainium.
It may deploy Fractile or other specialized inference hardware.
And, eventually, it could introduce increasingly Anthropic-specific silicon or custom accelerator designs.
That would give the company flexibility without forcing it to bet everything on one technology.
✅ Amir Salek’s Google TPU Background
The core claim is consistent with the supplied Bloomberg report: Salek is described as a veteran of Google’s custom-chip effort and former leader of the TPU business.
His background makes the appointment strategically relevant because TPU development is directly connected to the problem Anthropic is now trying to solve: large-scale AI compute efficiency.
✅ Anthropic Uses Multiple AI Hardware Platforms
Anthropic itself confirms that Claude is trained and operated across AWS Trainium, Google TPUs, and NVIDIA GPUs.
This supports the
✅ OpenAI Has Built Jalapeño
The comparison with OpenAI is factual.
OpenAI and Broadcom officially announced Jalapeño in June 2026 as an LLM-focused inference processor and described it as part of a multigenerational compute platform.
This strengthens the argument that custom silicon is becoming a strategic priority across frontier AI companies.
✅ Anthropic Has Expanded Google TPU Capacity
Anthropic announced a major Google and Broadcom agreement involving multiple gigawatts of next-generation TPU capacity expected to come online from 2027.
That means the
⚠️ Anthropic Is Not Yet Confirmed to Be Manufacturing Its Own Chip
This distinction is important.
The evidence supports Anthropic building expertise and exploring custom silicon, but that does not necessarily mean it has already become an independent semiconductor manufacturer.
The company could instead pursue co-design, custom accelerators, strategic investments, or partnerships with chip specialists.
⚠️ Fractile Deployment Is Still Future-Facing
The reported roughly $250 million Fractile agreement is significant, but the chips are not expected to be ready for use until 2027.
Therefore, this should be viewed as a forward-looking infrastructure commitment rather than current production capacity.
Prediction
(+1) Anthropic Will Become Much More Hardware-Oriented
Anthropic is likely to deepen its involvement in semiconductor design, accelerator architecture, and infrastructure engineering over the next several years.
The company does not necessarily need to manufacture chips itself to benefit from custom silicon.
(+1) Inference Chips Will Become Central to Claude’s Economics
As Claude usage grows and agentic workloads become more computationally intensive, inference efficiency will become increasingly important.
Specialized accelerators could eventually help Anthropic reduce cost and latency.
(+1) Anthropic Will Maintain a Multi-Chip Strategy
The company is unlikely to abandon NVIDIA, Google, or AWS.
Instead, Anthropic will probably continue mixing different processors based on workload, availability, price, and performance.
(+1) More Silicon Veterans Could Join Anthropic
The hiring of a senior Google TPU veteran could be an early signal that Anthropic intends to build a larger internal hardware organization.
If that happens, additional semiconductor architects, compiler engineers, systems designers, and infrastructure specialists could follow.
(+1) The AI Industry Will Become More Vertically Integrated
OpenAI’s Jalapeño and Anthropic’s expanding hardware strategy point toward the same future.
The leading AI companies increasingly want control over the entire stack—from models and software to processors and data centers.
(+1) The Real Winner Could Be Efficiency
The next major AI breakthrough may not simply be a model that scores higher on a benchmark.
It could be a model that delivers comparable or superior intelligence using dramatically less compute.
That would change the economics of the entire industry.
(-1) Custom Silicon Could Become an Expensive Distraction
There is also a significant downside.
If Anthropic invests heavily in proprietary hardware before model architectures stabilize, it could end up with expensive technology that becomes obsolete faster than expected.
(-1) Software Compatibility Could Become a Bottleneck
A custom processor is only valuable if Anthropic can build the software ecosystem required to exploit it.
If compiler, kernel, runtime, or developer tooling lags behind, theoretical hardware advantages may never become real-world advantages.
(+1) The Long-Term Direction Still Looks Clear
Despite those risks, the strategic direction is difficult to ignore.
AI companies are discovering that owning or influencing more of the infrastructure stack can provide greater control over cost, capacity, performance, and supply.
Amir Salek’s arrival therefore represents more than a senior executive hire.
It is another sign that the AI race is entering its next and far more physical phase—one where models, chips, data centers, electricity, and software must evolve together.
The next great AI advantage may not be hidden inside a model’s weights.
It may be sitting on a circuit board, drawing power in a data center, quietly generating millions of tokens while the rest of the world watches the chatbot on the screen.
🕵️📝Let’s dive deep and fact‑check.
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
References:
Reported By: www.deccanchronicle.com
Extra Source Hub (Possible Sources for article):
https://www.reddit.com/r/AskReddit
Wikipedia
OpenAi & Undercode AI
Image Source:
Unsplash
Undercode AI DI v2
🔐JOIN OUR CYBER WORLD [ CVE News • HackMonitor • UndercodeNews ]
📢 Follow UndercodeNews & Stay Tuned:
𝕏 formerly Twitter 🐦 | @ Threads | 🔗 Linkedin | 🦋BlueSky | 🐘Mastodon | 📺Youtube
