OCP Global Summit 2026 | Google | Partha Ranganathan | Intelligent Infrastructure for the Intelligence Era

1. From Rocket Engines to Microscopes: Google Changed the Subject

Last year, on the same stage, Partha Ranganathan's metaphor was that AI researchers are space explorers discovering new worlds and the OCP community builds the rocket engines. He called on the industry to go beyond ability to agility, building agile and fungible data centers so the same facility can take accelerators of different generations and vendors at any time.

This year's talk came in two halves. The first was a scorecard: over the past year Google and more than 50 OCP partners converged on a blueprint for fungible data centers. The second changed the subject: not just building systems for AI, but using AI to design systems. The closing metaphor moved from rockets to microscopes: if AI is deciphering the code of life, infrastructure builders are making the high-precision microscopes.

The first half tells you what Google is handing to the industry; the second tells you where Google intends to keep its differentiation.

2. The Vertical Integration Scorecard: Eight TPU Generations, Palomar OCS, 10-100x Gains

Google's argument starts from full-stack vertical integration: co-design all the way from energy, data centers, hardware and software to services and models. The "Radical AI System Innovations" slide lines up eight hardware innovations:

• Chips and memory: TPUs (now in their eighth generation, per Partha), Axion custom CPUs and GPUs; 288GB of co-packaged HBM.

• Power and cooling: through-substrate BGA vertical power, Deschutes CDU liquid-cooling loops, and 100 kW+ 3D pods for densification.

• Networking: the multi-TB/s ICI torus fabric and Palomar MEMS OCS optical circuit switching.

• Data centers: first-party and third-party (1P and 3P) gigawatt campuses.

▲ Google's eight AI system innovations: TPU/Axion/GPU, 288GB HBM, vertical power delivery, Deschutes liquid cooling, ICI torus, Palomar MEMS OCS, 100 kW+ 3D pods, and GW-scale campuses (Image source: Google | 2026 OCP Global Summit presentation (slide 6))

This vertical integration delivered 10x to 100x system efficiency gains, defined on the slide as performance, power and cost across generations, with no breakdown of which layer contributed how much.

What STT readers should note here is the optical network. OCS is the backbone of Google's clusters, and we covered how concentrated that market is in ECOC 2026 | 95% of OCS shipments are one company's business: Google. This keynote listed Palomar as one of the eight innovations, but with no open plan and no new numbers. The optical network remains the layer Google has no intention of sharing.

3. Breaking Down Power: From Diablo 400V to 800 VDC, Then to the Grid

Power was the densest layer of announcements. Partha described making data centers "DC-native from the grid to the chip", and two slides list the contributions and their co-sponsors clearly:

Item	Content	Co-sponsors
Diablo 400V	Disaggregated power rack	Google, Meta, Microsoft
LVDC distribution	Low-voltage DC power distribution spec	39 contributors, from ABB to Vertiv
Solid-state transformer (SST)	Medium voltage directly to 800 VDC	Google, Microsoft, NVIDIA
Grid interface	Performance under grid disturbances	Google, Meta, Microsoft, NVIDIA
Clean backup power	Backup technologies replacing diesel	Google, Microsoft

▲ Google's other OCP contributions: grid disturbance interface, SST from medium voltage to 800 VDC, clean backup power, LVDC distribution (39 contributors), the Open Data Center for AI white paper, and energy, carbon and water guidelines (Image source: Google | 2026 OCP Global Summit presentation (slide 12))

At the grid end, Partha added verbally that Google has about 1 GW of demand response under contract with U.S. utilities: during grid peaks, ML workloads can be shifted or reduced in exchange for faster access to power. The slide only says "demand response with utilities/grids", with no number.

Put together, the message is clear: the power bottleneck has moved from inside the rack to the moment of grid interconnection. When the four largest AI buyers sit down to write the same grid interface spec, getting power first matters more than having the fastest chip.

4. Breaking Down Cooling and Networking: CDUs, Liquid Retrofits, Protocol Interop

Cooling has two items:

• Project Deschutes: Google's multi-megawatt open coolant distribution unit (CDU). Partha said the number of suppliers adopting Deschutes has risen sharply, with many visible on the show floor.

• Project Brazos: bringing liquid cooling into existing air-cooled (brownfield) data centers; this was mentioned verbally only and is not on the slides.

▲ Three representative Google open contributions: Diablo 400V disaggregated power (Google, Meta, Microsoft), Project Deschutes CDU (Google), and Falcon networking (Google, NVIDIA) (Image source: Google | 2026 OCP Global Summit presentation (slide 10))

On the data center blueprint, the Open Data Center for AI white paper is signed by OCP, AMD, Google, Meta, Microsoft, NVIDIA and Oracle, defining common interfaces between server halls, IT and facilities.

Networking carries the most symbolic weight. A few years ago Google contributed its hardware transport protocol Falcon to OCP, with multipath load balancing, congestion control and line-rate encryption. This time Partha announced that NVIDIA, from BlueField to ConnectX, now has wire-format interoperability with Google's NICs, and the slide lists Falcon as a Google and NVIDIA effort. A rival's NIC running Google's protocol is one of the few cases where an open contribution has genuinely been adopted.

5. AI for Systems: Breaking Down Four Cases by the Numbers

In the second half, Partha used four cases to show how AI is rewriting system design, from the chip all the way to data center operations:

Layer	Case	Results published by Google
Systems hardware	AlphaEvolve LLM-guided evolutionary search to discover new computer architectures	Slide: new ideas and championship-winning benefits; spoken: 3-5x engineering velocity
System software	Mantis agentic security check and bug fix (open-sourced)	Every change scanned before merge, 92% precision, seconds of scan time
Data center design	AI-enhanced extreme-weather prediction models	Site-specific decadal, annual and weekly modeling; methodology opened
Operations and delivery	Automating drudgery and toil, capacity delivery	No public numbers yet

▲ The four levels at which Google uses AI to rewrite system design and the results it has disclosed; the closer to operations, the less quantified data there is (Image source: Simple Tech Trend original; data from Google 2026 OCP Global Summit presentation and keynote)

There were also several shorter examples: automatically translating TensorFlow to JAX, converting code between instruction sets, translating C++ to Rust, and using AI for scheduling, congestion management and power management.

Our view: the third case matters most for the hardware supply chain. A data center's cooling capacity is fixed at construction and, as Partha said, used for 15 years. The model combines physics-based simulation, historical weather and AI regressors to right-size infrastructure cost against downtime risk, cutting cooling equipment bought "just in case". If the method is widely adopted, cooling specs will shift from "design for the worst case" to "engineer per site", which could lower per-site volumes while raising the degree of customization.

Partha closed with a framework: AI transformation passes through three stages, impressive, useful and sustainable, while adoption moves in four steps: replicate (AI lift-and-shift), recombine (AI-augmented), reimagine (AI-first) and finally reinvent (AI-native). His analogy was the Ford Model T: early cars looked like horse-drawn carriages. By that framework, most of Google's own cases are still at AI-augmented.

6. Open Plumbing, Closed TPU and OCS

Put the two halves together and Google's division of labor becomes clear:

▲ Google contributes the grid interface, power, cooling, data center blueprints and transport protocol to OCP, while TPU, ICI, Palomar OCS and its models remain proprietary (Image source: Simple Tech Trend original; data from Google 2026 OCP Global Summit presentation)

• Handed to the industry: the grid interface, the 400V-to-800 VDC power chain, CDUs and liquid retrofits, the data center blueprint, and the transport protocol. These are all items where Google needs many outside suppliers and where more suppliers means lower prices.

• Kept in-house: TPU, the ICI interconnect, Palomar OCS, the Gemini models, and the AI-driven system design toolchain from the second half.

This is not a criticism. The business logic of the fungible data center is to turn electromechanical infrastructure into a commodity and maximize buyer leverage, while differentiation stays in compute and design speed. It matches our judgment in ECOC 2026 | No technology revolution in the next decade, only a manufacturing war: once specs are unified, competition shifts from technology to delivery speed and cost.

7. Risks and Counterarguments

• Most AI-for-systems numbers lack a public benchmark: the 92% precision is backed by an official blog post, but the 3-5x engineering velocity was only spoken, and the actual baselines in the paper (arXiv 2602.22425) still need checking.

• SST and LVDC are still at the spec stage: in the grid-to-chip DC-native architecture, SST has three co-sponsors and LVDC has 39 contributors, but the slides give no volume deployment timeline.

• 1 GW of demand response assumes deferrable load: training jobs can be scheduled, inference services much less so. The higher the inference share, the smaller this flexibility; and the number does not appear on the slides.

• No openness signal at all on the optical network: Palomar OCS is listed as a core innovation, but anyone hoping Google would bring OCS or an optical spec into OCP will be disappointed. The OCS supply chain still answers to a single customer, Google.

8. What It Means for Taiwan's Supply Chain

• 800 VDC and CDUs are the half being standardized. Diablo, LVDC and Deschutes are, or are becoming, OCP specs, so suppliers can build to spec without first winning a Google-private design. LVDC distribution alone has 39 contributors; Taiwanese vendors including Delta Electronics, LITEON and Auras also presented on 800 VDC and liquid cooling at this summit, showing they are actively positioning on this path.

• Standardization also means price competition. Partha said the number of Deschutes suppliers has risen sharply, which brings scale and margin pressure at the same time. Getting involved early in specs still taking shape, such as SST and clean backup power, will decide who ends up in the better position.

• Liquid retrofits (Brazos) are a separate market. Existing air-cooled facilities far outnumber new liquid-cooled builds; if retrofit demand is standardized, the serviceable market for thermal and mechanical vendors will expand.

• On the TPU and OCS half, Taiwanese vendors' position depends on Google's sourcing decisions, not on OCP specs. This session offered no new information there, so we draw no inference.

9. Conclusion

The most important takeaway from this keynote is Google's dividing line. Power, cooling, transport protocols and the data center blueprint are handed to OCP one by one to turn electromechanical infrastructure into a commodity; TPU, ICI, Palomar OCS and the AI design tools stay in Google's hands, and that is where the differentiation lies.

The timing needs to be read separately. Diablo and Deschutes are already OCP contributions, LVDC distribution has 39 contributors, and SST and the grid interface spec are still driven by a handful of companies. AI for systems is still at the "AI-augmented" stage of Google's own framework. Over the next 12 months, watch three things: when the SST spec is finalized, whether utilities formally adopt the grid interface spec, and whether Google's architecture search gets a public, reproducible benchmark.

Google is handing the plumbing to the industry and keeping the microscope for itself.

This article is for technology and industry trend analysis only and does not constitute investment advice.