Why Your AI Bill Is About to Get a Lot More Interesting
Powered by Single Fin
Welcome to this week’s edition of CIOsurge!
This week:
Why Your AI Bill Is About to Get a Lot More Interesting
How AI Is Really Changing the SOC
Forrester: AI and agents are redrawing the software market CIOs buy from
Let’s make this week a game-changer.
Stay sharp. Stay ahead.
Why Your AI Bill Is About to Get a Lot More Interesting
Every AI infrastructure conference ends with the same advice: buy more GPUs. Cerebras Supernova 2026 raised a more useful question, which is whether that spend is actually buying you anything. For a lot of workloads, the answer is no.
Here's the underlying mechanic. Every AI response is really two jobs. First the model reads and processes whatever you gave it, that's prefill, a raw compute problem. Then it generates the answer one token at a time, that's decode, a speed problem. Most companies run both steps on the same GPU without asking whether that's the right way to pay for either one. It's like hiring one person to both stock the warehouse and drive the delivery truck, then wondering why both jobs are slower and more expensive than they should be.
Split the job, cut the bill
Disaggregation means separating those two jobs and paying for each one separately. Prefill runs on AMD Helios, where raw throughput is what you're buying. Decode runs on Cerebras, where speed is what you're buying. The result is you stop overpaying for compute you don't need where you need speed, and stop underpaying for speed where you need throughput. For a company running high query volume, that's a direct reduction in token cost per response, not a marginal efficiency gain.
Credit to Andrew Feldman and Julie Choi for a session that treated this as an operating decision instead of a keynote pitch.
Cerebras also walked through CS-4 and the Nexus rack-scale platform: up to 30x faster than GPU on frontier models, 6x higher system performance than CS-3, a roadmap toward 10-trillion-parameter models at 1,000+ tokens per second. Treat the exact multiples with healthy skepticism until you test them, but the direction matters, the hardware is built to make split-workload inference cheaper at scale, not just faster in a demo.
Where this actually pays off
Figma's Pavi Bhattar showed the clearest business case. Their Design Agent is fine-tuned specifically to edit Figma files, trained on a decade of Figma's own design data. That's not a speed story, it's a cost and adoption story. A smaller, purpose-built model running fast enough to stay in the workflow costs less to run and gets used more than a bigger general model people have to wait on.
That's the real outcome here. People pay for time everywhere else in their lives, Uber, Disney Fast Pass, same-day shipping, and the same logic applies inside your company. A slow AI tool doesn't just cost more per token, it gets abandoned after the first try no matter how capable it is, which means the token spend you already made produced zero business value.
For CIOs, the move is to stop treating AI infrastructure as one undifferentiated GPU line item. Ask where prefill happens, where decode happens, and what it costs to separate them. That question is turning into one of the more reliable ways to lower AI spend while improving the response times that determine whether anyone keeps using the tool.
— Zack
How AI Is Really Changing the SOC
Security leaders have heard the pitch a hundred times: "AI will transform your SOC." The reality on the ground is messier.
When I sat down with Alexander Ondrick, Senior Manager of IT Security Operations, his view was clear: AI is not an add-on for the SOC. It is forcing a paradigm shift in how blue teams think about data, workflows, and people.
More queries, not better answers
Alex's team is experimenting with agentic workflows that will increase their query volume by 10x to 100x. That sounds great on a slide, but it exposes a hard truth: if the data underneath is broken, more AI just means more bad answers and more cost.
So his starting point for any AI initiative is not the model. It is a backward-planning exercise. What is the desired endstate, less toil or faster triage? What happens right before that, and where are the humans already in the loop? And do we have clean data and a solid asset inventory, or are we still in garbage-in, garbage-out territory?
His example is vulnerability management. If IPs keep changing via DHCP and the tool shows the same host three different ways, no amount of AI enrichment fixes that confusion. The unglamorous work of partnering with network and infrastructure teams to clean it up is what makes AI worth using.
Where humans stay in the loop
It is tempting to say that if AI can spot something, it should fix it. But patching windows, fragile dependencies, and systems that cannot go down do not care how smart your agents are. Alex's view is that you unlock nothing by defaulting to either extreme. Let automation handle scanning, correlation, and enrichment. Keep humans on prioritization and risk decisions.
The SOC is a team sport now
The old siloed model, with IR here and VM there and CTI somewhere else, is gone. As Alex put it, "that paradigm is dead. It's deader than dead." His SOC runs as a tight partnership across security operations, vulnerability management, threat intelligence, and detection engineering, sharing live channels so each team steers the others in near real time.
He also flagged the next blind spot: containers. Endpoint agents can tell him something is running inside a container, not what. Without that visibility, governing AI agents at scale is severely diminished.
— Alex & Zack
The Software Market You Buy From Is Being Redrawn
The software industry faced upheaval this year amid fears that AI and autonomous agents would replace traditional vendors, and CIO Dive reports that CIOs will need to think more long-term than the initial financial impacts suggest. Forrester research released this month found that advancements in AI will either replace or make current features of software platforms less important, according to Forrester VP and Principal Analyst Craig Le Clair. In some cases, frontier models are already taking over content generation and synthesis work that dedicated platforms were built to do.
Every renewal you sign in the next 18 months is a bet on which layer holds value. The uncomfortable question is which of your current platforms are essentially wrappers around capabilities the frontier models now do natively, because those are the ones whose pricing power is about to collapse. Pull your renewal calendar and sort it by that question. Shorten terms where the answer is unclear, and resist multi-year lock-ins on any tool whose core feature is something a general model already does well. The vendors know this is coming, which is exactly why they want longer contracts now.
— Zack Tembi
💡 CIO Spotlights
Guillermo Veiga named Chief Information Officer at Capital Group
Appointed CIO of Capital Group, the world's largest global active investment manager with $3.6 trillion in assets under management across 34 offices
Joins in November, relocating to California from Singapore, where he serves as Group Chief Information and Operations Officer at Standard Chartered Bank
Succeeds Marta Zarraga, who retires at the end of the year
Born in Uruguay and raised in Spain, began his career as a hands-on technologist and has held senior leadership roles across Europe and Asia at Amazon Web Services, Cisco, and Banco Santander
On the move, in his words: Capital is on the cutting edge of technology, and leading during global expansion as data and AI transform ways of working is an exciting prospect
Prabash Coswatte named Chief Information Officer at Trader Joe's
Named CIO of Trader Joe's, the specialty grocery chain
Joins from Heritage Grocers Group, where he held several senior roles including chief administrative officer and most recently COO, overseeing store operations and information technology
Served as CIO of Cardenas Markets beginning in 2019 during his Heritage tenure
Earlier was CIO at Vallarta Supermarkets and VP of enterprise program management at 99 Cents Only Stores
Steps in as Trader Joe's plans to open more than 20 locations in 2026, following 43 stores in 2025 and 34 in 2024
