번역 업데이트 중

Industry NewsGOOGL•9분 읽기

Google’s “Frozen v2” (Gemini-aware) chip targets 6–10× better tokens-per-watt by 2028—reshaping the AI inference hardware stack

Reuters/The Information reports Google is developing an internally named “Frozen v2” server chip that bakes Gemini model elements into hardware, targeted for as early as 2028 deployment. The chip is expected to deliver 6–10× more AI tokens per unit of power than Google’s latest custom silicon and is intended to complement (not replace) Google’s existing TPU roadmap—aiming to relieve compute bottlenecks as AI capex rises. For investors, the key question isn’t only whether the chip works, but whether Google can turn improved tokens-per-watt into measurable inference cost leverage versus competitors’ GPUs/accelerators, with TSMC likely central to the advanced packaging and manufacturing ramp.

게시일 2026년 7월 21일업데이트 2026년 7월 21일

Reported project name

Frozen v2

Internal server chip (informally dubbed), reported by Reuters citing The Information

Target deployment

As early as 2028

Engineering/design still being finalized per report snippets

Efficiency goal

6–10×

AI tokens served per unit of power vs. Google’s latest custom chips

Role vs. TPUs

Complement

Designed to complement, not replace, the existing TPU line

Google’s AI chip strategy just moved from “better accelerators” to “model-aware silicon.” The reported “Frozen v2” roadmap targets a step-change in inference efficiency (6–10× tokens-per-watt) and pushes deployment toward 2028, signaling that the next battle in AI infrastructure may be won by those who reduce power per token—not just compute FLOPs.

Reported project name

Frozen v2

Internal server chip (informally dubbed), reported by Reuters citing The Information

Target deployment

As early as 2028

Engineering/design still being finalized per report snippets

Efficiency goal

6–10×

AI tokens served per unit of power vs. Google’s latest custom chips

Role vs. TPUs

Complement

Designed to complement, not replace, the existing TPU line

What happened (and what’s actually new)

Google isn’t only optimizing Gemini software—it’s trying to hardwire Gemini into inference hardware

The load-bearing novelty is the “Gemini-aware hardware” idea: Reuters (citing The Information) reports “Frozen v2” aims to incorporate elements of the Gemini AI model directly into chip architecture. That’s a different optimization target than general-purpose accelerator throughput—because it can reduce energy wasted on generic execution paths.

Frozen v2: key claims to anchor your expectations

Chip

Frozen v2 (internal server chip)

Reported by Reuters citing The Information

Deployment timing

Targeted for 2028

As early as 2028 per reporting

Metric

Tokens per unit of power

Efficiency measured as AI tokens served per power

Expected improvement

6–10×

Relative to Google’s latest custom AI chips

Strategic positioning

Complement TPUs

Not a replacement; part of a broader compute-stack control push

Important: the report describes projected efficiency and intent; it does not (yet) provide silicon test results or a public benchmark methodology. Treat 6–10× as a target until Google later discloses comparable, auditable performance.

The mechanism

How Gemini-aware silicon can realistically produce 6–10× tokens-per-watt (and where the risk hides)

A 6–10× jump in tokens-per-watt is only plausible if “Frozen v2” reduces energy in at least two places at once: (1) fewer cycles per token through model-structured execution, and (2) lower memory movement and control overhead via tighter hardware-software coupling. The risk is that gains on a subset of workloads may not generalize to the full Gemini inference mix.

Tokens-per-watt is dominated by energy per effective token, which typically includes compute + memory + orchestration. Model-aware hardware can target the orchestration and dataflow parts, not only math throughput.
If chip architecture incorporates “elements” of Gemini, it can reduce wasted execution on generic paths (e.g., specialized decode/update flows, structured attention handling, or hardwired routing).
Complementing TPUs suggests a heterogeneous strategy: Frozen v2 could take the “most power-hungry” inference slices, while TPUs cover flexibility or other model families—limiting software-migration risk.

What must be true for a 6–10× tokens-per-watt target to hold up in practice
Requirement	What the claim implies	Why it matters for investors	What to watch next
Comparable benchmark scope	Efficiency measured on representative Gemini inference workloads	Otherwise, 6–10× could be an apples-to-oranges lab result	Later disclosures: workload mix, batch sizes, sequence lengths, KV-cache handling assumptions
Memory-movement reduction	Lower energy spent on moving weights/activations and managing KV cache	Because memory bandwidth/latency often limits real tokens-per-watt	Signs of advanced on-package/high-bandwidth memory integration and system-level design
Thermal/power headroom at scale	Tokens-per-watt maintained under datacenter cooling and power delivery constraints	Datacenter power caps can erase “bench” advantages	Any system-level power envelope specs, rack-level density statements
Software maturity	Compilers/runtime able to exploit hardware model hooks reliably	Without tooling, hardware targets can’t be realized	Early adoption in Google services and performance regression stability

Fundamentals + capacity context (where this matters inside Alphabet)

Alphabet has the cash generation to fund an internal compute stack—but execution still decides the ROI

Even without Frozen v2 operating metrics, the financial setup matters: Alphabet shows strong operating cash generation and substantial ongoing capex capacity. The investment question becomes whether improved inference efficiency translates into lower cost per token (and thus more margin or more capacity for demand) rather than simply shifting the constraint from chips to power delivery, networking, or utilization.

Alphabet (TTM) operating cash flow

$174.4B

Net cash provided by operating activities (TTM snapshot)

Alphabet (TTM) capex

$109.9B

Investments in property, plant & equipment (TTM)

Alphabet (TTM) free cash flow

$64.4B

Free cash flow (TTM)

Balance sheet cushion (TTM)

$380.6B

Total assets (TTM) and $126.8B cash + short-term investments

Alphabet cash capacity: operating cash flow and free cash flow (FY 2023–TTM 2026)

Use this to frame whether Alphabet can self-fund multi-year compute refresh cycles while sustaining AI capex.

단위: USD

Alphabet Operating cash flow (FY 2023)

Net cash provided by operating activities

101,746,000,000

Alphabet Operating cash flow (FY 2024)

Net cash provided by operating activities

125,299,000,000

Alphabet Operating cash flow (FY 2025)

Net cash provided by operating activities

164,713,000,000

Alphabet Operating cash flow (TTM 2026)

Net cash provided by operating activities

174,353,000,000

Alphabet Free cash flow (FY 2023)

Free cash flow

69,495,000,000

Alphabet Free cash flow (FY 2024)

Free cash flow

72,764,000,000

Alphabet Free cash flow (FY 2025)

Free cash flow

73,266,000,000

Alphabet Free cash flow (TTM 2026)

Free cash flow

64,429,000,000

Cash capacity helps, but doesn’t guarantee returns: if utilization stays low (or if system-level bottlenecks appear), tokens-per-watt improvements may not translate into proportionate unit economics.

Supply chain + who benefits/loses

This pushes demand toward advanced packaging and datacenter integration—not just leading-edge wafers

A Gemini-aware server chip is still a server platform, which means the supply chain story is bigger than “one chip design.” If Frozen v2 truly targets 6–10× better tokens-per-watt, it likely increases the importance of memory bandwidth, power delivery, and packaging integration—areas where TSMC is positioned as a critical manufacturing/packaging partner for advanced AI compute platforms.

Likely supply-chain linkage from model-aware inference silicon to public markets
Layer	What changes with Frozen v2	Why it matters for efficiency	Example public beneficiaries (tickers resolved)
Wafer fabrication	More advanced process nodes or specialized silicon options for performance-per-watt	Directly impacts transistor efficiency and high-frequency behavior	TSMC
Advanced packaging / system integration	Higher bandwidth memory and tighter thermal/power design in a server form factor	Real tokens-per-watt is often memory + packaging energy, not only compute	TSMC
Datacenter power + thermal infrastructure	Potential rack density changes and power delivery design for higher efficiency chips	If power delivery caps are reached, chips can’t express their efficiency	Not modeled here (no additional tickers in the brief)
Server OEM / integration	Custom boards/cooling/power rails to match Frozen v2’s power profile	System-level integration determines real tokens-per-watt in deployed racks	Not modeled here (no additional tickers in the brief)

Upstream: If TSMC supports Frozen v2 wafer + packaging steps, then any design shift that improves bandwidth or thermal density can increase value capture in advanced packaging volumes.
Downstream (customers of compute): Google Cloud inference consumers may see lower cost per query if utilization and pricing enable it—otherwise the benefit may mostly accrue to Google’s internal cost structure.
Competitive implication: the chip is designed to complement TPUs, meaning Google can tune deployment by workload class rather than forcing a single architecture across all inference.

Competitive positioning

The “Nvidia alternative” thesis depends on system-level economics, not just tokens-per-watt in a vacuum

If Frozen v2 delivers the targeted energy efficiency at scale, it can undercut competing inference stacks that rely on general-purpose accelerators. But investors should avoid the trap of assuming better chips automatically win: real economics depend on how much capacity Google can actually sell or allocate, how quickly it can migrate workloads, and whether power/cooling and networking bottlenecks dominate.

What to measure (or wait for) to validate whether Frozen v2 changes competitive dynamics
Signal	What would confirm it	What would falsify it	Why it matters
Latency + throughput parity	Google can serve Gemini workloads with equal or better end-to-end latency under the same power cap	Efficiency gains come with unacceptable latency or throttling	Cloud buyers optimize for SLA and user experience, not only energy
Utilization improvements	System deployment raises effective tokens-per-watt at rack level	Compute stays underutilized due to scheduling or workload mismatch	Unutilized accelerators turn theoretical gains into idle power costs
Pricing/unit-economics impact	Lower cost per token feeds into either margin expansion or more capacity for demand	Savings are absorbed internally without monetization or capacity release	Determines shareholder impact, not just engineering success
Heterogeneous deployment effectiveness	Frozen v2 complements TPU usage with a clear workload partition	Architecture overlap creates internal complexity and lowers overall efficiency	Complementing TPUs only helps if the division of labor is coherent

Investable view (1–3 year horizon)

Google’s next equity catalyst is whether tokens-per-watt becomes a repeatable, monetizable cost advantage

Frozen v2 is a forward-looking project (2028 target), so the near-term catalyst is less about immediate financial numbers and more about direction of travel: AI compute bottlenecks, datacenter capex efficiency, and any early internal deployment learnings that reduce uncertainty about realization.

Base case (data-driven): If Google sustains strong cash generation (operating cash flow) while funding capex, it can absorb schedule and performance risks long enough to reach 2028.
Upside case: Model-aware silicon reduces cost per token meaningfully and measurably; then internal compute economics can support either lower prices or higher capacity for demand growth.
Downside case: Bench efficiency fails to translate to system-level tokens-per-watt due to memory/network/thermal constraints or workload mismatch—making the chip a niche improvement rather than an inflection.

Why this matters now: the report’s specificity (Frozen v2, 6–10× tokens-per-watt target, 2028 deployment, complementing TPUs) gives investors a concrete benchmark to track against future disclosures.

About Us•Privacy Policy•Terms of Service•Join our Slack