Confidential compute comes to Akash. 5 Min. Read Jul 28, 2026 Decentralized compute has always faced an underlying question that requires a direct, transparent answer rather than a reassuring platitude: what prevents a completely unknown provider from accessing your data when you rent their GPU? Previously, the solution relied on provider reputation, security audits, and the assumption that operators lacked the time or incentive to inspect individual containers. While acceptable for minor tasks like hosting a game server, this baseline falls short when protecting sensitive medical records, proprietary algorithmic trading models, or core company intellectual property embedded in model weights. Akash is introducing a definitive solution today with the launch of confidential compute across the network. Workloads can now run inside hardware-enforced trusted execution environments (TEEs). In this setup, memory encryption is managed directly by the CPU, rendering the data inaccessible to the host operating system, the hypervisor, and anyone with physical access to the infrastructure. The most notable aspect of this feature is the minimal effort required to implement it. One line. Here is the entire SDL change: Integrating confidential computing requires no SDKs, rebuilt container images, separate orchestrators, or new deployment types. You simply utilize params.tee, which configures cpu for CPU-isolated workloads or cpu-gpu when an enclave-protected GPU is necessary. The marketplace handles matching seamlessly behind the scenes. The SDL builder translates the designated tee value into a placement requirement, ensuring only providers verifying confidential compute capability bid on your order. Rather than manually selecting a specific TEE vendor, the hardware infrastructure dictates whether you receive Intel TDX or AMD SEV-SNP, while provider software automatically provisions the appropriate runtime class. Once a lease is finalized, your container runs isolated within a Kata Containers micro-VM, utilizing a dedicated kernel and encrypted memory instead of sharing host kernel resources. If your configuration requests a GPU, pass-through is managed via VFIO into NVIDIA Confidential Computing mode, safeguarding all data moving across the PCIe bus with AES-GCM-256 encryption. Developed publicly under AEP-83 and inheriting foundations from AEP-65 and AEP-29, this architecture maintains a deliberately constrained SDL surface. By prioritizing the reuse of existing frameworks and minimizing new additions, confidential computing remains strictly opt-in, preserving current deployments entirely without disruption. A private LLM in one deployment. To demonstrate how this abstraction works in practice without overstating its simplicity, a complete, functional example called ollama-tee has been added to the awesome-akash repository. It runs Ollama with Llama 3.2 3B on a GPU, inside a TEE: Note the image: it's the stock ollama/ollama:latest from Docker Hub. Nothing in it knows or cares that it's running in an enclave. Deploy it, wait for the model to pull, and query it the way you'd query any Ollama instance: An ordinary API response, but with one crucial difference: the weights, prompt, and completion were all handled within memory that remains completely hidden from the machine's operator. This encapsulates the entire value proposition of confidential computing, explaining why the SDL modifications are restricted to a single line. Most people will simply avoid privacy measures if they demand a complete application rewrite. What this protects, and what it doesn't. Let's be exact about its security guarantees. Hardware Protection Capabilities: The CPU's memory encryption engine encrypts guest memory, keeping it hidden from the host OS, the hypervisor, or root operators on the machine. It also encrypts CPU-to-GPU communication via the PCIe bus, turns off GPU performance counters to eliminate side channels, and runs every workload with a dedicated kernel inside its own micro-VM. Limitations of TEE Configuration Alone: Simply requesting a TEE does not confirm that you are operating within a genuine, uncompromised environment, as isolation is distinct from proof. Verification requires remote attestation, which gathers signed data from within the VM (such as an SEV-SNP report or TDX quote, alongside GPU data from NVIDIA's NVTrust) to validate it against baseline measurements and generate a token for relying parties. In setups combining CPU and GPU TEEs, this becomes composite attestation, where the CPU TEE serves as the primary trust anchor for the GPU. While the ollama-tee demonstration highlights isolation and placement, it does not perform attestation. Because attestation is mandatory for regulated applications, Akash Network want to be straightforward instead of overstating TEE capabilities. Akash Network is currently developing tooling to integrate attestation as a smooth, integrated component of the deployment lifecycle. Akash Network should also note two technical details regarding Hopper GPUs: CPU-to-GPU data transfers within a TEE are restricted to roughly 4 GB/s due to a software bounce buffer, which slows down model loading rather than active inference. This bottleneck is resolved in Blackwell through PCIe IDE and TDISP. Additionally, NVLink communication between Hopper GPUs lacks encryption, which is important when sharding models, but Blackwell introduces NVLink encryption to fix this. Why this matters now. Historically, the exact workloads that stood to benefit the most from decentralized infrastructure were completely blocked from adopting it. For example, a medical network reviewing an AI diagnostics system cannot expose confidential health records to an unaudited machine. Similarly, engineering groups managing high-value proprietary model weights cannot risk deploying them to an unidentified operator. Furthermore, any autonomous agent managing client PII must fully comply with the privacy mandates bound to that data. The core obstacle for these operations has never been about infrastructure costs or hardware constraints, especially since Akash has continuously delivered compute resources at a fraction of traditional hyperscaler pricing. Instead, the real limitation stemmed from compliance; a policy requiring teams to simply "trust the operator" cannot serve as a verifiable security control within formal regulatory documentation. By introducing hardware-enforced confidentiality, reliance on operational trust is substituted with a far more narrow and verifiable premise: that the physical silicon executes precisely as specified by the manufacturer. This shift fundamentally alters the security dynamic, finally making decentralized compute accessible for sensitive, enterprise-grade workloads that require absolute data privacy. Get started. * Deploy the example: ollama-tee in awesome-akash * Read the spec: AEP-83 * Deploy from the browser: Akash Console * Bring questions: Akash Discord Because confidential compute remains opt-in and provider capacity is in its early stages, you might encounter initial challenges - such as receiving no bids, facing SDL validation errors, or discovering attestation questions left unanswered by the documentation. If you run into these obstacles, please reach out to Akash Network on Discord. The specification is completely open, the core implementation uses the Apache 2.0 license, and user feedback at this critical phase directly influences the final design.

Crypto Race 2026 - In the bear market, staking is starting to make a difference. July 18, 2026 checkpoint. The Crypto Race 2026 is still moving through a difficult phase. With the exception of Akash, every asset in Category 1 is now showing a negative return since the race began on January 3. In Category 2, only Gnosis remains ahead of Bitcoin. The main takeaway from this checkpoint comes from the comparison between the two rankings tracked since January: * without rewards, Bitcoin moves back into third place; * with rewards included, ATOM remains on the podium. After almost seven months, staking and node yields are therefore starting to affect the ranking in a concrete way. CAT 1 - Akash stumbles, Bitcoin moves up without accelerating. Akash remains well ahead, but loses another 9.6% over the week. Its performance since the beginning of the race falls back to +40.9%. The lead remains substantial: AKT still holds more than a 60-point advantage over Bittensor. However, the leader is no longer extending the gap and is gradually giving back part of its gains. Bittensor remains in second place despite a weekly decline of 8.8%. Its lead over Bitcoin is now limited to seven points. Bitcoin returns to the podium. Bitcoin ends the week almost unchanged, with a weekly move of −0.2%. That relative stability is enough to overtake ATOM, which fell by 5%, and reclaim third place in the ranking excluding rewards. BTC now shows a performance of −29.1%, compared with −30.2% for Cosmos. Only 1.1 points separate the two assets. Ethereum is the only CAT 1 asset to post a positive week, rising by 2.5%. ETH nevertheless remains seventh since the start of the race. Avalanche continues to sit at the bottom of the ranking with a performance of −52.6%. CAT 2 - Gnosis strengthens its position, Nolus drops back. Gnosis remains clearly in control of CAT 2. With a performance of −15.5%, GNO is still the only asset in this category outperforming Bitcoin. Its weekly gain of 5.4% strengthens that position further. Band Protocol moves up one place and returns to second. FLUX and Shentu also climb in the ranking despite posting negative weekly performances. Their rise is mainly explained by another sharp fall from Nolus. Nolus loses three places. Nolus falls by 30.2% in one week. Its performance since the start of the race drops to −66.5%. NLS moves directly from second to fifth place, overtaken by Band, FLUX, and Shentu. After spending a long period among the more resilient assets in CAT 2, Nolus has lost most of that advantage in just two weeks. Lava records the strongest weekly performance in the category, rising by 9.5%, but remains last with an overall decline of 88.2%. The main ranking is based on the price performance of each asset. Based on price alone, Bitcoin leads ATOM by 1.1 points. The ranking changes once the rewards accumulated since January are included. With rewards, ATOM keeps its podium position. The estimated rewards improve ATOM's performance by around six points. Cosmos therefore remains third ahead of Bitcoin in the ranking that includes staking. The same difference can be seen in CAT 2: The overall order remains unchanged, but the difference is significant for Band and Shentu. The specific case of FLUX. For FLUX, this is a node yield rather than staking. The current annual yield varies approximately between 37% and 47%, depending on the node tier. Based on price alone, FLUX currently shows a performance of −60.2%. Once an estimated node yield is included, the overall performance would be approximately: * −51.5% with an annual yield close to 37%; * −48.8% with an annual yield close to 47%. The result remains negative, but the gap compared with the token-only performance reaches almost ten points. Conclusion. Akash remains at the top of the Crypto Race 2026 despite another significant decline. Bittensor stays second. Bitcoin returns to third place in the ranking excluding rewards, ahead of ATOM. In CAT 2, Gnosis continues to dominate while Nolus falls by three positions. The main point of this checkpoint remains the difference between the two calculation methods. Without rewards, Bitcoin is third. With rewards included, ATOM keeps its podium position. After 196 days of racing, staking and node yields do not offset major price declines, but they are beginning to affect final performance and, in some cases, the ranking itself.