Google Unveils TPU 8tx and 8i Max: Specialized AI Chips for Secure and Long-Context Workloads

Google Expands Eighth-Generation AI Chip Portfolio

Google has unveiled two specialized Tensor Processing Units—TPU 8tx and 8i Max—expanding its eighth-generation AI chip lineup to four systems in a single year1

1

3

3

. Amin Vahdat, Google's chief technologist for AI infrastructure, confirmed these are reconfigured versions of existing silicon designed to address specific customer demands rather than entirely new chip designs2

2

. The move signals Google's strategic shift toward specialized AI chips that can be deployed rapidly—in months rather than years—by adjusting packaging, memory, and networking around foundational chip architectures3

3

.

These new variants join the TPU 8t and TPU 8i, which Google introduced at Cloud Next in April for large-scale training and real-time reasoning respectively1

1

. The TPU 8t and 8i deliver up to 2.7 times better training performance and 80% better inference price performance compared to previous generations3

3

. Google's accelerated release schedule marks a departure from its previous cadence of launching one new processor every two years1

1

.

TPU 8tx Brings Hardware-Level Encryption for Regulated Industries

The TPU 8tx targets secure and confidential AI workloads for government agencies and companies in regulated industries1

1

4

4

. Built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts, the 8tx implements confidential computing by keeping sensitive data and model weights encrypted even during active processing3

3

. Normally, data must be decrypted in memory while computers work on it, creating security vulnerabilities2

2

.

The chip encrypts data at the hardware layer through a secure enclave, eliminating the need to expose information during processing4

4

. By adopting the same low-latency Boardfly network created for real-time reasoning, TPU 8tx delivers up to 2x lower latency compared to TPU 8t with near-zero performance penalty3

3

. This combination of training-grade compute with fast inference networking allows organizations to simultaneously run and customize models on a single confidential system, enabling regulated services to fine-tune models using private data without migrating between separate clusters4

4

.

TPU 8i Max Scales Memory for Massive Context Windows

The TPU 8i Max addresses the growing demand for long-context AI applications by delivering double the compute, double the SRAM, and double the high-bandwidth memory of TPU 8i4

4

. Vahdat emphasized this matters particularly for drug discovery, mathematics, and coding applications that require processing vast amounts of information in a single request2

2

. The chip packs two TPU chips together, providing extra processing muscle and memory to handle data without constant swapping2

2

.

Google developed the 8i Max alongside pushing massive context windows for its Gemini frontier model, which supports the 1 million-token industry standard—roughly equivalent to 750,000 words or 1,500 pages of text4

4

. The system fits twice the silicon in the same rack footprint, achieving the highest compute density per rack while lowering floor space requirements per square foot and per watt in data centers3

3

4

4

. Adding more on-chip memory reduces the need to move data between processors, lowering latency and reducing power consumption as AI models process information4

4

.

Positioning Against Nvidia in the Agentic Era

Google's in-house chips have emerged as credible alternatives to Nvidia's graphics processors, which dominate AI computing1

1

5

5

. The company rents out Tensor Processing Units through its cloud business rather than selling them as standalone chips1

1

. Anthropic holds the right to acquire up to $150 billion of TPUs from 2028 through 2031, according to regulatory filings2

2

. Google's TPU customer base has grown to nearly 100 organizations, including frontier labs, trading firms, retailers, and universities3

3

.

Vahdat explained the strategic rationale: "We could build a single chip that tries to address the whole market, or by building specialized versions of the chips, we actually can deliver much more value"2

2

. The new variants are positioned specifically for the agentic era, as businesses deploy autonomous agents that carry out tasks on users' behalf while facing growing security risks1

1

. Google has also announced plans to make TPUs available through Oracle Cloud Infrastructure in 2027, marking the first time the chips will be offered on a hyperscale cloud platform beyond Google's own infrastructure2

2

.

Supply Constraints and Manufacturing Strategy

Demand for Google AI chips remains intense. Every TPU unit Google produces is already committed to a customer, according to the company's earnings reports2

2

. Google faces shortages of memory and manufacturing capacity and is actively working to diversify its suppliers2

2

. The company currently works with four design partners, including Broadcom and MediaTek, to build its Tensor Processing Units, with manufacturing handled by TSMC1

1

2

2

. Vahdat declined to comment on the specific roles of Marvell Technology or MediaTek in the TPU business2

2

. The new TPU 8tx and 8i Max systems are expected to become available next year2

2

.