The Puget Sound's Hydroelectric Edge Brought Databricks to Bellevue — Now the Hyperscalers Are Gobbling the Last of It
The Bellevue Bet: 270,000 Square Feet of AI-Native Engineering
Databricks signed a 160,000-square-foot lease at Four106, a 21-story tower at 380 106th Ave. NE in downtown Bellevue, in June 2026. The deal pushed the company's total Puget Sound footprint past 270,000 square feet, more than double its prior regional count and a figure that dwarfs its earlier presence here. The Puget Sound Business Journal first reported the lease, which covers a majority of the available office space in the 485,000-square-foot building developed by Patrinely Group and Dune Real Estate Partners and completed in 2025.
The Four106 lease did not come out of nowhere. Databricks had been assembling its Bellevue position piece by piece: it occupied space at City Center Bellevue in increments, including a 37,500-square-foot expansion in October 2025, and added roughly 18,000 square feet across five floors in a 27-story tower. Earlier in 2026, the company was already in talks with Uber to lease a combined 420,000 square feet at Four106. Uber ultimately signed for approximately 170,000 square feet in April, and Anthropic also committed to space in the building. The result is a Bellevue cluster where three frontier-AI companies occupy the same structure, a concentration that does not exist elsewhere on the West Coast outside San Francisco.
The timing matters. New office construction in downtown Bellevue has slowed, and Four106 is one of the few recent deliveries with large blocks of available space. The building sits roughly half a mile from Bellevue's downtown light rail station, a practical detail for a company hiring across the region. Zero G Talent's board shows 55 Databricks roles added in the past week alone, spanning Berlin, Munich, Seoul, Belgrade, London, Mexico City, and São Paulo — but the Puget Sound campus is where the company's core data-engineering and ML-operations teams sit.
The Bellevue footprint now puts the Puget Sound campus in the same conversation as the company's largest Bay Area holdings, including 150,000 square feet in San Francisco and 305,000 square feet in Sunnyvale. For a company with more than 7,000 employees worldwide and 20 global offices, the Puget Sound expansion is not a satellite. It is a primary node, and its scale tells you what kind of work happens there.
Why Former AI Chief Ali Ghodsi's Post-Databricks Bet Matters
Naveen Rao ran Databricks' AI division during the company's push to build the dominant enterprise data-and-AI platform. Now he's betting that the real bottleneck isn't models or data — it's the power bill. His startup, Unconventional AI, claims it can cut AI inference energy use by a factor of 1,000, and it has already demonstrated a working image-generation model, Un-0, running on a software simulation of its oscillator-based chip architecture. The company released schematics for actual hardware and plans to build a full inference stack from the ground up.
The claim is staggering, and the logic behind it is straightforward. "AI scaling is hard because of energy. It's going to be the fundamental limit in the next few years," Rao told TechCrunch. "You just can't go past it." If he's even directionally right, the implication is seismic: the geography of AI infrastructure stops being determined by fiber routes and tax incentives and starts being determined by who has cheap, abundant, zero-carbon power.
That's what connects Rao's departure to Databricks' Bellevue expansion. At the Data + AI Summit 2025, Databricks CEO Ali Ghodsi laid out a "Data Intelligence for All" roadmap (Agent Bricks, Lakebase, LakeFlow Designer, MLflow 3.0) all built on the premise that AI agents will need to run on enterprise data at scale, in production, under human supervision. The workloads those roles support will consume real electricity, in real data centers, in real time.
Rao's bet is that the current silicon stack — GPUs built on conventional digital logic — will hit a wall that no amount of packaging advances can fix. His oscillator-based architecture is designed to perform certain classes of inference computation using a fundamentally different physical process, one that he says could reduce power consumption by three orders of magnitude. The current Un-0 demo runs on a software simulation; the hardware is still forthcoming. But if the company delivers even a fraction of that efficiency gain, it would reshape the economics of where inference workloads land.
The Pacific Northwest already has the power profile this thesis demands. The region's hydroelectric infrastructure, dams on the Columbia and Snake rivers, produces some of the cheapest electricity in the United States, and PNNL is actively developing digital twin technology to optimize hydropower operations for data center loads. A building design standard discussed at the Northwest Power Conservation Council could reduce peak energy consumption in data center electrical and mechanical systems by one-fifth. Rao doesn't need to build in Washington to benefit from this logic, but the same logic is what makes Databricks' Bellevue campus more than a real estate play.
The energy-efficiency race isn't theoretical. OpenAI unveiled its first custom chip, built by Broadcom, the same week Rao went public with Un-0. The entire industry is looking for ways to decouple AI scaling from power scaling. Rao's approach is the most radical (a new computing paradigm rather than an optimization of the old one) and the most uncertain. But if it works, the companies and regions sitting on cheap hydro power won't just have a cost advantage. They'll have the only kind of advantage that matters once AI hits the energy ceiling.
The Lakebase Rollout and the Enterprise-Data Pivot
Databricks' Lakebase product, its managed Postgres-compatible database service, is quietly becoming the connective tissue between the company's AI platform ambitions and the enterprise data infrastructure that already exists inside Fortune 500 companies. The service handles transactional workloads and analytical queries in a single engine, which means the data engineering teams enterprises already employ can start running AI-adjacent pipelines without re-architecting their stack. That matters for workforce demand because it collapses two job categories, the traditional data engineer who manages ETL pipelines and the ML operations engineer who serves models, into a single hybrid role. Databricks listed a Lakebase Sales Specialist opening in both Mexico City and São Paulo in the past week alone, signaling that the product is in active global rollout rather than beta.
The Nokia partnership, announced in June 2026, gave that workforce thesis its first concrete telecom proof of concept. The joint PoC validated a unified, substrate-agnostic data platform designed to support AI-driven autonomous networks. This system lets operators ingest network telemetry at Tier-1 scale and run the same data pipelines across Databricks and open-source environments built on Apache Flink, Kafka, and Iceberg without rewriting code. Nokia engineers built a platform-independent data transformation layer in Python that separates business logic from infrastructure connectors. A custom compiler then translates those abstract workflows into native execution formats, Delta Live Tables on one end and Flink SQL on the other, at deployment time. The architecture supports query-time data products, zero-copy sharing across operational domains, and multi-agent AI systems that perform root-cause analysis and cross-domain correlation autonomously.
"Telecom operators are managing increasingly complex networks and need a more consistent way to harness their data. Our collaboration with Nokia demonstrates how a unified data platform can help simplify operations and unlock the value of AI across network domains." — Nevash Pillay, Global Head of Telecommunications Industry, Databricks
What makes this relevant to the Puget Sound workforce story is the skill set it demands. The Nokia-Databricks PoC requires engineers who understand both streaming data infrastructure and AI agent orchestration. These are people who can write the Python transformation logic, configure the Kafka ingestion layers, and then plug the output into an agentic AI system that generates new data products from natural-language prompts. That's not a pure software engineering role. It's not a pure data science role. It's the hybrid that Databricks' entire Lakebase-and-Agent Bricks product line is designed to create demand for.
Nokia, for its part, is pushing the same integration from the network side. At DTW Ignite in Copenhagen, the company expanded its autonomous networks portfolio with agentic AI capabilities spanning RAN, IP, fixed, and optical networks. The company's stated target is Level 4/5 autonomous network operations, systems that don't just automate tasks but make real-time decisions. Getting there requires exactly the kind of unified data layer the PoC with Databricks validated.
Databricks ranked highest in both Execution and Vision in the 2025 Gartner Magic Quadrant for Data Science and Machine Learning Platforms, its fourth consecutive year as a Leader. That enterprise credibility is what makes the Lakebase pivot work: CIOs who already trust Databricks with their analytics workloads can adopt Lakebase for transactional use cases without a separate vendor evaluation. The data engineers they already employ become the AI-data hybrid workforce Databricks needs at scale.
This is the demand signal behind the Bellevue square footage. Not model training. Not inference hosting. The harder, less glamorous work of building the data infrastructure that makes AI agents useful inside enterprises — and the workforce that infrastructure requires.
Berlin, Munich, Seoul, Bellevue: A Global Talent Corridor
The Bellevue campus is not a standalone bet. It is one node in a simultaneous, multi-continent hiring push that is stitching together a distributed but unified AI-intelligence workforce, and the job postings tell the story in real time.
Databricks is hiring solutions engineers in Munich and Berlin with the same job codes it uses for its U.S. roles. A Munich posting for a Solutions Engineer, EMEA Greenfield, lists the same Field Engineering Launchpad program, the same 20–30% travel requirement, and the same core skills: Python, SQL, Apache Spark, AWS or Azure. A nearly identical role in Berlin requires English and lists German as a plus, with the same emphasis on proofs-of-concept and cloud platform fluency.
The Seoul expansion is further along. Databricks opened its Korea office in 2022 and recorded over 100% year-on-year growth in its Korean business by April 2024, driven by enterprise demand for data and AI capabilities. The company is now hiring a Sr. Specialist Solutions Engineer (AI/ML) in Seoul, a role that pairs data science and machine learning engineering with customer-facing pre-sales work. It's the same hybrid profile showing up in Germany and Washington State.
What connects these postings is not just a brand. It is a single workforce architecture. Databricks is building a global layer of technical staff who can deploy the same Data Intelligence Platform against different regulatory regimes, energy markets, and enterprise maturity levels, and then share the resulting playbooks across offices. Bellevue gives the company access to the Pacific Northwest's hydroelectric power and a deep bench of cloud infrastructure talent. Munich and Berlin give it a foothold in the EU's strict-data-governance market. Seoul gives it a position inside one of the world's fastest-growing enterprise AI adoption curves.
The result is a talent corridor that runs from the Puget Sound to the Han River, with the same job codes at both ends.
Does Puget Sound Power Change the AI Infrastructure Map?
The Pacific Northwest spent a century selling its cheap hydroelectric power as its defining competitive advantage. Then the data center boom nearly drained the surplus. Now, as AI workloads send energy demand soaring past what the grid can reliably deliver, that old advantage is mutating into something stranger: a bottleneck that simultaneously attracts and constrains the companies building the next generation of frontier AI.
Washington state's electricity costs have climbed since 2024, and state regulators project further increases as the state's clean energy mandates collide with hyperscale data center demand. The Washington Department of Commerce's July 2025 draft findings flagged a blunt reality: the power system cannot add capacity on pace with demand growth, constrained by permitting timelines, interconnection backlogs, transmission limits, and the sheer physical lead time of building new generation at scale. Puget Sound Energy's vice president of clean energy strategy, Josh Jacobs, put it more directly at a November 2024 Tech Alliance panel in Seattle: "We are short today. The hyperscalers are growing today. They're gobbling up available hydro and carbon-emitting resources that are on the market today."
That shortage is the context for Databricks' Bellevue commitment. The company isn't just leasing office space; it's positioning itself inside a region where the single largest operating cost for any AI infrastructure buildout, electricity, is simultaneously cheap by national standards and under severe structural pressure. The Northwest Power and Conservation Council's Ninth Power Plan process, which kicked off formal work in February 2025 with adoption targeted for the end of 2026, is explicitly designed to forecast data center load growth over the next 20 years. The Council's December 2024 expert panel heard from Brian Janous, former Microsoft VP of Energy and co-founder of Cloverleaf Infrastructure, who warned that data center growth in the Northwest is likely to be constrained until 2030 by electricity availability, with grid-enhancing technologies and storage as the only near-term relief valves.
The hyperscalers are not waiting for the grid to catch up. Amazon is funding the feasibility phase for a 320-megawatt nuclear facility in central Washington. Microsoft signed a power purchase agreement with Helion, a Seattle-area fusion company that has not yet generated electricity from a commercial plant. Google already operates a data center near the Columbia River in The Dalles, Oregon, drawing on the same hydroelectric infrastructure. These are not sustainability gestures; they are hedges against a grid that cannot guarantee the firm, 24/7 clean power that AI training clusters require at scale.
For Databricks, the calculus is different from a hyperscaler's. The Bellevue campus is an engineering hub, not a GPU farm. The company's compute workloads run on cloud infrastructure and its own distributed systems, not on dedicated megawatt-scale clusters in Bellevue. But the location still matters. The Pacific Northwest's power constraints are pushing the geography of AI infrastructure toward a bifurcated model: training happens where firm power is available or contractually locked in, while engineering and R&D cluster where talent lives and energy costs remain manageable. Bellevue sits at the intersection of both logics. It offers access to the same hydroelectric grid that powers the region's data centers, proximity to Amazon and Microsoft's Pacific Northwest infrastructure footprint, and a talent pool that already spans the AI-data-engineering boundary the company's Lakebase product is designed to serve.
The risk is that the window closes. The Department of Commerce's draft findings warned that large data center loads could force regulators to socialize grid upgrade costs across existing ratepayers, or create stranded utility investments if hyperscalers exit before cost recovery. Janous framed it as a coordination problem: "There's so much capital that wants to invest in energy infrastructure. The problem you have is that there's not that many opportunities right now to invest that capital efficiently."
Databricks' Bellevue expansion is, in this light, a bet that the Pacific Northwest's energy constraints will resolve in the company's favor — or at least not resolve against it. The office space is a physical commitment to a region whose power economics are in flux. If the Ninth Power Plan process, the nuclear investments, and the grid-enhancing technology deployments materialize on anything close to schedule, the Puget Sound retains its cost advantage over AI hubs in Texas, Arizona, and Northern Virginia, where power is either more expensive, more carbon-intensive, or both. If they don't, the company has still hired the engineers. The hydroelectric moat is real, but it has a drawbridge, and the hyperscalers are racing to control it.
What 270,000 Square Feet Signals About Post-Training AI Demand
The Bellevue campus is not a satellite office. It is a physical thesis about where AI infrastructure work is headed. When a company with over 10,000 employees worldwide and a Glassdoor Best Place to Work 2025 nod commits that much square footage to a single metro, it is betting that the workforce it needs over the next decade cannot be assembled in San Francisco alone. More precisely, the kind of engineer it needs increasingly does not want to live there.
That workforce is hybrid by design. Databricks' open positions in Bellevue span distributed data systems engineering, AI/ML solutions architecture, and field engineering roles that sit between customers and the platform. A new-grad distributed data systems role in Bellevue carries a listed base of roughly $133,000 to $143,000, according to a Dice posting. A solutions architect role targeting AI/ML customers is listed on LinkedIn. These are not back-office seats. They are the roles that build, deploy, and maintain the data-intelligence layer that enterprise AI runs on.
The Lakebase product roadmap sharpens the picture. Lakebase branching, Databricks' approach to letting enterprises run isolated, governed copies of production data environments, demands engineers who understand both the storage substrate and the orchestration layer. It is a product that sells to regulated industries: financial services, healthcare, the Fortune 500 accounts Databricks says make up over 60% of its customer base. Those customers do not want a pure research scientist. They want someone who can stand up a compliant data environment, wire it into an existing on-premises network, and keep it running. The Nokia autonomous-network demonstrations point in the same direction: the next wave of AI deployment is infrastructure-heavy, operations-heavy, and location-flexible.
This is where the Puget Sound's hydroelectric advantage stops being an energy story and becomes a workforce story. Cheap, predictable power lowers the cost of running large-scale data and AI workloads. That makes Bellevue a cheaper place to operate a data-intelligence platform than Northern Virginia, where wholesale electricity prices have climbed, or Texas, where grid reliability remains a concern for hyperscale operators. Lower operating costs mean the campus can scale headcount without the same margin pressure that constrains Bay Area offices, where commercial real estate and compensation costs compound. The result is a structural incentive to grow the Bellevue headcount faster than the San Francisco one.
The global hiring pattern confirms it. Databricks is adding roles in Berlin, Munich, Seoul, and Bellevue simultaneously. Zero G Talent's own board shows a Lakebase Sales Specialist opening in Mexico City and another in São Paulo alongside the European and Asia-Pacific engineering hires. This is not a cost-arbitrage play, shipping jobs to cheaper markets. It is a distributed build: each region gets a piece of the platform, and Bellevue is the anchor for the data-systems and AI-infrastructure layer specifically. The company's own careers page lists over 30 offices in more than 20 countries, but the Bellevue campus is one of the few with a dedicated engineering footprint at this scale.
What emerges is a workforce built for the unglamorous reality that follows the model launch: the pipelines that break at 2 a.m., the compliance audits that demand on-site engineers, the enterprise customers who won't touch a platform that can't run inside their own walls. The Puget Sound was an unlikely home for this five years ago. With the new campus, a hydroelectric cost base, and a product roadmap that demands infrastructure engineers who can work on-site with regulated customers, it is becoming the default one.
Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at Databricks, and the people building the field.