Google unveils two specialized AI chips to challenge Nvidia
Google is launching the TPU 8tx and TPU 8i Max to provide custom hardware for high-security data and large-scale AI research.
The short version
- Google is launching the TPU 8tx for high-security tasks and the TPU 8i Max for expanded context windows.
- The TPU 8tx features an AMD processor to enable hardware-level data encryption.
- The TPU 8i Max combines two chips to increase memory and computing power for complex research.
- Google and Oracle will make these chips available on Oracle Cloud Infrastructure starting in 2027.
- Anthropic has an agreement to purchase up to $150 billion worth of TPUs between 2028 and 2031.
Why it matters
This shift shows that hardware for AI is becoming more specialized rather than one-size-fits-all. It also signals a new era of cloud cooperation where Google's hardware will run on a competitor's infrastructure.
A shift toward specialization
Google is introducing two new custom-designed tensor processing units, or TPUs, to compete more directly with Nvidia. These new chips, the TPU 8tx and the TPU 8i Max, are scheduled to launch next year. The move signals a departure from Google's previous two-year release cycle for its processors.
Amin Vahdat, who heads Google’s chip division, explained that building purpose-built hardware offers more benefits than creating a single universal chip for the entire market. Instead of trying to serve every possible use case with one design, Google is tailoring its hardware to meet specific customer needs. This strategy allows for more efficient performance in niche applications.
While Google's TPUs are designed to challenge Nvidia's GPUs, they have already secured major customers like Meta, Apple, and Oracle. The company continues to work with Broadcom on design and TSMC on manufacturing to bring these products to market. However, Google is currently facing supply chain constraints as its entire existing TPU inventory is already reserved.
Hardware-level security
The TPU 8tx is a specialized chip designed for tasks that require a high level of security. It is primarily aimed at government agencies and businesses operating in heavily regulated sectors. By focusing on these specific requirements, Google hopes to capture a market that demands more than standard AI processing.
A key technical feature of the TPU 8tx is the integration of an AMD processor. This allows the chip to perform data encryption directly at the hardware level. This is a significant improvement over current methods where data typically has to be decrypted in system memory during the processing phase.
By keeping data encrypted while it is being worked on, the chip reduces potential vulnerabilities. This hardware-level approach provides a more robust security layer for sensitive information. It addresses a common pain point for organizations that must handle private or restricted data sets.
Expanding AI context windows
The TPU 8i Max serves a different purpose by targeting AI capabilities that require expanded context windows. A context window is the amount of data a model can process in a single query. This capability is essential for complex fields such as mathematics, pharmaceutical research, and software engineering.
To achieve these high-performance goals, the 8i Max combines two separate TPU chips into a single unit. This configuration increases the total memory and computing power available to the user. It allows for the processing of much larger datasets or longer sequences of information than standard chips might allow.
These new products are not entirely original architectures from the ground up. Instead, they are modified versions of existing chips that have been optimized for specific customer needs. This allows Google to iterate faster while still providing the specialized power required for high-end research and development.
Cloud access and partnerships
In a significant move for the cloud computing landscape, Google and Oracle have announced a partnership regarding these new chips. Starting in 2027, TPUs will be accessible through Oracle Cloud Infrastructure. This marks the first time Google's chips will be available on a competitor's hyperscale cloud service.
This collaboration highlights the growing demand for specialized AI hardware across the industry. Even as competitors like Nvidia diversify their own offerings, the demand for diverse hardware options remains high. For example, Nvidia is developing its own CPUs and using Groq’s technology for specific inference hardware.
The scale of demand is already evident in recent regulatory filings. Anthropic has reached an agreement to purchase up to $150 billion worth of these units between 2028 and 2031. This massive investment underscores the importance of Google's hardware in the future of large-scale AI development.
Addressing supply chain hurdles
Despite the ambitious rollout and high demand, Google faces significant hurdles in production capacity and memory. Recent financial reports indicate that the current inventory of TPUs is completely spoken for. To keep up with the needs of its customers, Google intends to expand its supplier base.
While the company is looking to add more partners to the manufacturing process, Vahdat did not specify which companies would be involved. Potential candidates for these roles could include firms like Marvell Technology or MediaTek. Expanding the supply chain is necessary to ensure that the 2027 and 2028 launch dates remain viable.
The success of these new chips will depend on Google's ability to navigate these logistical challenges. By diversifying its manufacturing partners and focusing on specialized designs, Google aims to provide a viable alternative to Nvidia. The industry will be watching closely to see how these new units perform in real-world research and security environments.
What to watch
- Whether Google adds MediaTek or Marvell to its manufacturing supply chain.
- How the TPU 8tx performs in real-world government security audits.
- The performance of the 8i Max in pharmaceutical research compared to Nvidia's latest GPUs.
Try it
Check if your company's cloud provider offers access to specialized AI hardware like TPUs or GPUs for your specific data security needs.
Sources and links
First reported by techcentral.ie. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.