{"id":51516,"date":"2026-05-26T14:28:07","date_gmt":"2026-05-26T14:28:07","guid":{"rendered":"https:\/\/www.europesays.com\/ai\/51516\/"},"modified":"2026-05-26T14:28:07","modified_gmt":"2026-05-26T14:28:07","slug":"hybrid-by-design-engineering-the-new-model-for-federal-ai-delivery","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ai\/51516\/","title":{"rendered":"Hybrid by design: Engineering the new model for federal AI delivery"},"content":{"rendered":"

This is the 14th article in our IT lifecycle management series,\u00a0Delivering the tech that delivers for government<\/a>.<\/p>\n

Federal agencies have spent years balancing cloud and on\u2011premise environments. Now, artificial intelligence is forcing them \u2014 and the systems integrators that support them \u2014 to design for both, deliberately and at scale.<\/p>\n

As AI moves from pilot projects into daily operations, this new model is taking hold across government and its delivery partners. It is defined not by where systems run, but by how fast and securely they deliver outcomes, said Graham Gilmer, a senior vice president at Booz Allen whose work focuses on autonomy and AI in the defense sector<\/a>.<\/p>\n

Hybrid architecture is no longer a transitional phase. It is the foundation for delivering AI at mission speed, Gilmer said. \u201cDepending on the use case, you may want to go with one or both or a hybrid combination for redundancy.\u201d<\/p>\n

It\u2019s a fundamental change in mindset. As Mike Watkinson, chief revenue officer at Future Tech Enterprise<\/a>, put it, \u201cThe mission defines the architecture, and the architecture enables and secures the mission.\u201d<\/p>\n

That principle is not new. What is new is how directly it is shaping technical decisions \u2014 down to the level of compute form factor, data placement and engineering workflow.<\/p>\n

During a panel discussion for our series, Delivering the tech that delivers for government<\/a>, we asked Gilmer and Watkinson to share how this is setting new infrastructure realities for the government and its technology partners.<\/p>\n

From centralized systems to distributed delivery<\/p>\n

For more than a decade, cloud computing provided a path to scale. But it did not eliminate the need for control, latency management or operational resilience.<\/p>\n

Today\u2019s environments require systems that can operate across classification levels, across geographies and often without persistent connectivity, Gilmer said. That is driving agencies toward distributed architectures that deliberately combine cloud infrastructure, on\u2011prem environments and edge systems. Plus, there are new user expectations on how they should be able to access the tools that they need to do their jobs.<\/p>\n

\u201cWe are constantly monitoring the gap between commercial \u2014 what\u2019s available on your phone or outside of the skiff where you may work \u2014 to what\u2019s available within government,\u201d he said. \u201cI\u2019m pleased to say that that is actually closing.\u201d<\/p>\n

Both Gilmer and Watkinson said the emphasis is moving from standardization to orchestration, creating new expectations for systems integrators to design, integrate and operate across environments, not just deploy within a single stack.<\/p>\n

New Reality 1: GPU form factors are redefining what\u2019s possible<\/p>\n

One of the most important technical shifts enabling AI use in the hybrid model is the rapid evolution of GPU\u2011based compute.<\/p>\n

Eighteen months ago, running meaningful large language models required access to hyperscale infrastructure. Today, advances in graphic processing units are changing that.<\/p>\n

GPU memory density, power efficiency and packaging are enabling desktop\u2011class and even portable deployments, Watkinson pointed out. Systems with 100-plus gigabytes of VRAM can now support medium\u2011sized language models suitable for real mission workflows, he said.<\/p>\n

Watkinson described these systems as at an inflection point, noting that desktop GPUs \u201caccelerate the AI and accelerate the outcomes in the mission.\u201d They simplify deployment by collapsing compute, storage and security controls into a single enclosure, eliminating many of the integration challenges associated with distributed cloud environments, he said.<\/p>\n

Gilmer said there has been a fairly swift uptake in on-prem desktop use all the way to the edge to make AI tools accessible there. \u201cIt used to be only small, very small, LLMs,\u201d he said. \u201cNow, I\u2019d say it\u2019s up to medium, very capable LLMs \u2014 things that can change how missions are run, or how enterprises operate, or certainly sensitive or classified workloads are processed in the government.\u201d<\/p>\n

The takeaways? In practice, this means agencies can now:<\/p>\n

Run containerized models locally for inference
\nDeploy AI into classified or disconnected networks without external dependencies
\nStand up rapid prototyping environments without making multiyear infrastructure investments<\/p>\n

That means the constraint is no longer access to compute, Watkinson said. For integrators, that shifts the challenge from procuring infrastructure to integrating AI capabilities quickly into mission workflows.<\/p>\n

New Reality 2: Inference \u2014 not training \u2014 is shaping architecture<\/p>\n

As agencies scale AI, they are discovering that the dominant cost is not building models but running them.<\/p>\n

Inference workloads \u2014 including real\u2011time queries, agent interactions and automation pipelines to deliver results from models \u2014 now account for the majority of compute demand. \u201cIt\u2019s pivoted significantly to inference, 80% to 90%,\u201d Gilmer said, describing inference as a \u201chuge variable cost,\u201d driven by query volume and increasingly by agent\u2011based systems executing continuously.<\/p>\n

This change has direct architectural implications for how and where AI workloads run, particularly for integrators responsible for balancing cost, performance and deployment across environments.<\/p>\n

Cloud environments offer elasticity, but they also introduce data egress and ingress costs, latency variability and potentially less predictable pricing at scale. Localized GPU deployments, by contrast, enable more controlled cost models.<\/p>\n

\u201cThe cost of iterating and the cost of experimentation is less,\u201d Watkinson said.<\/p>\n

He pointed to lower \u201ccost per token\u201d and reduced iteration costs when workloads run closer to the data. This is particularly important as agencies test and refine AI applications, where iteration speed is as critical as steady-state performance.<\/p>\n

The result is a more nuanced deployment strategy, both Gilmer and Watkinson said. Training may still occur in centralized environments. But inference is increasingly distributed \u2014 on\u2011prem, at the edge or in hybrid configurations.<\/p>\n

New Reality 3: Data locality and control drive design decisions<\/p>\n

While compute is becoming more flexible, data still requires tight governance, which continues to be a persistent challenge for agencies with vast data stores.<\/p>\n

That makes data strategy a shared responsibility between agencies and their integration partners, especially in environments with strict classification boundaries.<\/p>\n

\u201cIt\u2019s not only how much data, but which data am I working with,\u201d Watkinson said, underscoring a challenge that is becoming central to AI architecture. Federal agencies must balance accessibility with strict requirements around classification, privacy and provenance.<\/p>\n

Gilmer added that in many cases, \u201cthere are reasons we wouldn\u2019t want data to leave a certain area,\u201d which is driving the need for on\u2011prem and localized processing.<\/p>\n

That reality is reshaping system design. AI pipelines must now account for data residency requirements, version control across environments and access controls tied to identity and role (as defined by federal zero trust frameworks).<\/p>\n

In practical terms, this often means bringing the AI model to the data rather than the data to the model, a reversal of earlier cloud\u2011centric assumptions.<\/p>\n

New Reality 4: Engineering workflows are evolving with AI<\/p>\n

The impact of these architectural changes extends directly into engineering practices.<\/p>\n

The impact is not just technical. It changes how engineering itself is done \u2014 especially for integrators delivering and maintaining systems on behalf of agencies. AI is becoming an active participant in development.<\/p>\n

Gilmer described how \u201cwe\u2019re now significantly more productive,\u201d citing measurable gains in software engineering output. Development environments are increasingly supported by AI agents that can generate code, refactor legacy systems and assist with testing.<\/p>\n

This is particularly impactful in federal environments, where legacy codebases present ongoing challenges.<\/p>\n

AI systems can help:<\/p>\n

Translate outdated languages into modern frameworks
\nExtract business logic from legacy applications
\nAccelerate modernization without requiring scarce expertise<\/p>\n

\u201cHaving a development agent work alongside human engineers \u2014 we\u2019re going to see this shift continue,\u201d Gilmer said. The result is a new engineering model that blends human judgment with machine\u2011driven scale.<\/p>\n

New Reality 5: Speed continues to redefine delivery<\/p>\n

All of these changes converge on a single outcome: faster delivery.<\/p>\n

Traditional acquisition cycles and development timelines no longer align with the pace of AI innovation. Working together, agencies and their technology partners are adopting a model built on rapid iteration and real\u2011world validation.<\/p>\n

\u201cWe have to field to learn,\u201d Gilmer said, which requires deploying early and refining continuously. Watkinson added that this enables teams to \u201citerate faster and experiment more,\u201d lowering the barrier to trying new approaches. That\u2019s happening across the government technology space, both in agencies and at their contractors.<\/p>\n

Speed is no longer a secondary metric. It is a primary measure of mission success. Watkinson put it succinctly, \u201cHow much do you value speed to market?\u201d is now a defining question for agencies investing in AI.<\/p>\n

Key design principles for hybrid AI<\/p>\n

Start with mission requirements, not infrastructure
\nDesign for inference cost, performance and scalability
\nKeep data local when control or classification requires it
\nDesign for integration across environments and vendors
\nEnsure operation across cloud, on\u2011prem and edge environments
\nPrioritize speed, iteration and continuous improvement<\/p>\n

Discover more smart tips and tactic for FSIs shared by leading technologists for our Delivering the Tech That Delivers for Government series<\/a><\/p>\n

Copyright
\n \u00a9\u00a02026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.\n <\/p>\n","protected":false},"excerpt":{"rendered":"This is the 14th article in our IT lifecycle management series,\u00a0Delivering the tech that delivers for government. Federal…\n","protected":false},"author":2,"featured_media":51517,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[24,25,30491,28507,30492,2676,30493,30494,30495],"class_list":["post-51516","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ai","tag-ai","tag-artificial-intelligence","tag-booz-allen","tag-booz-allen-hamilton","tag-delivering-the-tech-that-delivers-for-government","tag-future-tech","tag-future-tech-enterprise","tag-graham-gilmer","tag-mike-watkinson"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/51516","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=51516"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/51516\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/51517"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=51516"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=51516"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=51516"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}