A machine with 6,100 qubits can make 1,024 simultaneous attempts to crack your encryption for every 10 qubits it runs. Do the math, and the locked vault your security team spent years building starts looking less like a fortress and more like a screen door.
That is the uncomfortable reality facing enterprise IT right now. Quantum computing is no longer a research novelty tucked away in university basements. According to MIT's latest Quantum Index Report, more than two dozen manufacturers are commercially selling over 40 quantum processing units today. The quantum-as-a-service model is spreading access even further, meaning you do not need to own a quantum computer to use one.
To be clear, quantum machines are not yet running laps around classical computers on real-world business problems. MIT's own researchers acknowledged that current quantum processors still fall short of what is needed for large-scale commercial applications like cryptanalysis or complex chemical simulations. IBM's Institute for Business Value echoed that, noting we have not reached the tipping point where quantum consistently outpaces traditional computing.
But here is the part that should keep security leaders up at night: the threat does not have to be fully mature to be dangerous. Adversaries are already harvesting encrypted data today with the explicit plan to decrypt it later, once quantum capability catches up. By the time quantum is ready, your 2025 data could be wide open.
Tim Steward, a principal data enterprise architect at Fujitsu, put it plainly at a recent industry conference. The assumption that encryption equals safety is no longer a safe assumption. The confidence most organizations have placed in 128-bit encryption is built on a foundation that quantum computing is actively working to erode.
The National Institute of Standards and Technology has already issued its recommendation: migrate to 256-bit encryption as soon as operationally possible. That single move could buy enterprises meaningful protection for the next two decades, giving security teams room to breathe while the quantum landscape continues to evolve.
Early enterprise quantum use cases are still focused on pharmaceuticals and molecular modeling, with supply chain optimization expected to follow. These applications sound distant from your security posture, but they signal that investment in quantum is accelerating fast, which means the timeline for cryptographic risk is compressing alongside it.
The uncomfortable truth is that quantum security is not a future problem being handed to a future team. The decisions made in the next 12 to 18 months about encryption standards, data retention policies, and vendor partnerships will determine how exposed organizations are when quantum capability crosses the threshold that actually matters. Waiting for that threshold to arrive before acting is precisely the wrong strategy.
AI
Pinterest Slashed AI Costs by 90 Percent With One Bold Engineering Move
Pinterest did not negotiate better rates with an AI vendor or find a cheaper cloud region. It cut its AI inference costs by 90 percent by doing something most engineering teams would consider borderline reckless: it surgically removed the vision layer from a frontier model and rebuilt what it actually needed from scratch.
The result is a case study in what happens when a company stops treating AI models like black boxes and starts treating them like engineering problems with parts that can be taken apart and reassembled.
Frontier models are designed to be general-purpose. That versatility is exactly what makes them expensive. When you run a massive multimodal model to do one specific job, you are paying for enormous capability you are not using. Pinterest, a platform built almost entirely around visual discovery, realized that a significant portion of what it was running through its AI pipeline did not need the full stack. It needed a narrow, precise capability executed very cheaply at massive scale.
The engineering team's approach was to strip out the vision processing layer that comes bundled with the frontier model and replace it with a purpose-built alternative tuned specifically for Pinterest's use case. This is not fine-tuning in the conventional sense. It is closer to performing open-heart surgery on a model and replacing a major organ with something more efficient for the specific body it is living in.
The 90 percent cost reduction is striking, but the more important number is what that savings unlocks. AI inference costs are one of the primary reasons companies throttle how aggressively they deploy machine learning features in production. When you cut those costs by nearly an order of magnitude, you can run more experiments, serve more users, and build features that were previously too expensive to justify.
This matters well beyond Pinterest's engineering org. The broader AI industry has spent two years debating whether frontier models are worth their price tags. Pinterest's move is a concrete argument that for many production use cases, the answer is: not if you engineer around them carefully.
It also puts pressure on the model providers themselves. If customers can carve out the expensive parts and get 90 percent of the value at 10 percent of the cost, the pricing leverage that frontier model companies currently enjoy starts to erode. Expect more engineering teams to study exactly what Pinterest did here and attempt to replicate it across different verticals.
The playbook is not easy. It requires deep technical expertise, a well-defined use case, and the willingness to diverge from the path of least resistance. But for companies running AI at scale, the potential return on that engineering investment just got a very compelling proof point.
Source: VentureBeat
The result is a case study in what happens when a company stops treating AI models like black boxes and starts treating them like engineering problems with parts that can be taken apart and reassembled.
Frontier models are designed to be general-purpose. That versatility is exactly what makes them expensive. When you run a massive multimodal model to do one specific job, you are paying for enormous capability you are not using. Pinterest, a platform built almost entirely around visual discovery, realized that a significant portion of what it was running through its AI pipeline did not need the full stack. It needed a narrow, precise capability executed very cheaply at massive scale.
The engineering team's approach was to strip out the vision processing layer that comes bundled with the frontier model and replace it with a purpose-built alternative tuned specifically for Pinterest's use case. This is not fine-tuning in the conventional sense. It is closer to performing open-heart surgery on a model and replacing a major organ with something more efficient for the specific body it is living in.
The 90 percent cost reduction is striking, but the more important number is what that savings unlocks. AI inference costs are one of the primary reasons companies throttle how aggressively they deploy machine learning features in production. When you cut those costs by nearly an order of magnitude, you can run more experiments, serve more users, and build features that were previously too expensive to justify.
This matters well beyond Pinterest's engineering org. The broader AI industry has spent two years debating whether frontier models are worth their price tags. Pinterest's move is a concrete argument that for many production use cases, the answer is: not if you engineer around them carefully.
It also puts pressure on the model providers themselves. If customers can carve out the expensive parts and get 90 percent of the value at 10 percent of the cost, the pricing leverage that frontier model companies currently enjoy starts to erode. Expect more engineering teams to study exactly what Pinterest did here and attempt to replicate it across different verticals.
The playbook is not easy. It requires deep technical expertise, a well-defined use case, and the willingness to diverge from the path of least resistance. But for companies running AI at scale, the potential return on that engineering investment just got a very compelling proof point.
Enjoyed this?
Get stories like this delivered every Tuesday — free.